A power equipment defect single task identification analysis method and system

Abstract

The application belongs to the technical field of power equipment defect analysis, and discloses a power equipment defect single task identification analysis method and system. The power equipment defect single task identification analysis method comprises the following steps: a task processing complexity comprehensive index is constructed to evaluate the complexity of the power equipment defect single task data to be identified, and the task processing complexity is obtained; based on the task processing complexity, an integrated analysis framework is constructed to adaptively allocate a differential learning device, and a target identification model corresponding to the power equipment defect single task data to be identified is determined; according to the task processing complexity and the target identification model, collaborative reasoning scheduling is performed between an edge node and a cloud center, and the computing power of the target identification model is adaptively allocated in combination with node computing power, network bandwidth and task execution delay, so as to output an identification analysis result. The model collaborative efficiency, real-time performance and adaptability in the power equipment defect single task identification analysis process are improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment defect analysis technology, and in particular to a single-task identification and analysis method and system for power equipment defects. Background Technology

[0002] As the scale of power systems continues to expand, power equipment in transmission, substation, and distribution links operates in complex environments for extended periods. Defects such as equipment surface damage, insulation abnormalities, overheating, and abnormal vibration, if not detected promptly, can easily affect operational safety. To improve inspection efficiency and defect identification accuracy, multi-source data such as images, soundprints, vibration, and temperature are increasingly being applied to power equipment condition sensing and defect analysis. Artificial intelligence-based defect identification has also become an important technological direction in intelligent operation and maintenance of power equipment.

[0003] In existing technologies, for complex recognition tasks, models of different sizes are typically configured based on task requirements and resource conditions to balance recognition performance and operating costs. Small models are characterized by low computational cost, fast response speed, and flexible deployment, making them suitable for rapid judgment in resource-constrained or real-time-critical scenarios. Large models have strong feature extraction and representation capabilities, making them suitable for tasks with large datasets, complex feature relationships, or high recognition difficulty. By coordinating and scheduling models of different sizes, the advantages of different models can be leveraged to some extent, improving the overall processing efficiency of the system and reducing the resource consumption caused by the continuous operation of a single large model.

[0004] However, existing technologies still have certain shortcomings in power equipment defect analysis scenarios. On the one hand, power equipment defect analysis often requires the integration of multimodal data such as images, sound patterns, vibration, and temperature, while existing collaborative scheduling frameworks are mostly designed for single-modal or relatively fixed data formats, making it difficult to adapt to the complex sources of power equipment defect data. On the other hand, the operating status of power equipment is continuous and real-time, and defect identification and analysis need to provide judgment results as soon as possible. Existing frameworks still have bottlenecks in real-time inference efficiency, making it difficult to fully meet the power system's requirements for rapid response. Furthermore, existing collaborative scheduling methods mostly rely on static preset rules, lacking the ability to adapt to dynamic changes in defect identification and analysis tasks. Different defect types have significantly different requirements for model complexity and analysis accuracy. If the model collaboration method cannot be flexibly adjusted according to real-time data characteristics, it can easily lead to redundant computing resources or a decrease in identification accuracy.

[0005] Therefore, how to improve the efficiency, real-time performance, and adaptability of model collaboration in the single-task defect identification and analysis process of power equipment is a problem that needs to be solved by existing technologies. Summary of the Invention

[0006] This invention provides a method and system for single-task identification and analysis of defects in power equipment, in order to solve the problems of poor multimodal data adaptability, low real-time inference efficiency and inflexible model scheduling in the existing single-task identification and analysis of defects in power equipment.

[0007] According to a first aspect of the present invention, a single-task identification and analysis method for defects in power equipment is provided.

[0008] In one embodiment, the single-task identification and analysis method for power equipment defects includes: assessing the complexity of the single-task data of power equipment defects to be identified by constructing a comprehensive task processing complexity index; the comprehensive task processing complexity index includes information complexity index, task complexity index, and process complexity index; based on the task processing complexity, adaptively allocating differentiated learners by constructing an integrated analysis framework to determine the target identification model corresponding to the single-task data of power equipment defects to be identified; the differentiated learners include random forest model, extreme gradient boosting tree model, and one-dimensional convolutional residual network model; according to the task processing complexity and the target identification model, performing collaborative inference scheduling between edge nodes and the cloud center, and adaptively allocating the computing power of the target identification model by combining node computing power, network bandwidth, and task execution latency to output the identification and analysis results.

[0009] According to a second aspect of the present invention, a single-task defect identification and analysis system for power equipment is provided.

[0010] In one embodiment, the power equipment defect single-task identification and analysis system includes: a complexity assessment module, used to assess the complexity of the power equipment defect single-task data to be identified by constructing a comprehensive task processing complexity index, thereby obtaining the task processing complexity; the comprehensive task processing complexity index includes information complexity index, task complexity index, and process complexity index; a model allocation module, used to adaptively allocate differentiated learners based on the task processing complexity by constructing an integrated analysis framework, thereby determining the target identification model corresponding to the power equipment defect single-task data to be identified; the differentiated learners include a random forest model, an extreme gradient boosting tree model, and a one-dimensional convolutional residual network model; and a collaborative scheduling module, used to perform collaborative inference scheduling between edge nodes and the cloud center according to the task processing complexity and the target identification model, and adaptively allocate the computing power of the target identification model by combining node computing power, network bandwidth, and task execution latency, so as to output the identification and analysis results.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] (1) This invention constructs a comprehensive index of task processing complexity to quantitatively evaluate the single task data of power equipment defects to be identified from the aspects of information complexity, task complexity and process complexity. It can accurately reflect the processing difficulty of different defect identification tasks, provide a basis for subsequent model selection and task scheduling, and thus avoid the inefficiency caused by using the same processing method for tasks with different complexities.

[0013] (2) This invention constructs an integrated analysis framework and adaptively allocates between the random forest model, the extreme gradient boosting tree model and the one-dimensional convolutional residual network model, so that the single task of power equipment defects with different complexities can be matched with the corresponding target recognition model. It can reduce the computational resource consumption of low-complexity tasks by using lightweight models, and ensure the recognition accuracy of high-complexity tasks by using complex models, thereby realizing targeted learning and recognition decision optimization of multi-source fault features.

[0014] (3) This invention performs collaborative reasoning scheduling between edge nodes and cloud centers, and adaptively allocates computing power based on task processing complexity, node computing power, network bandwidth and task execution delay. It can select appropriate reasoning execution paths according to real-time resource status, reduce unnecessary computational redundancy and transmission overhead, reduce task processing time and energy consumption, and improve the real-time performance, stability and robustness of single-task identification and analysis of power equipment defects in different operating scenarios.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] Figure 1 This is a specific embodiment diagram shown according to an exemplary embodiment;

[0018] Figure 2 This is a confusion matrix diagram of the 1DCNN-Ret algorithm illustrated according to an exemplary embodiment;

[0019] Figure 3 This is a confusion matrix diagram of the XGBoost algorithm according to an exemplary embodiment;

[0020] Figure 4 This is a confusion matrix diagram of the random forest algorithm according to an exemplary embodiment;

[0021] Figure 5 This is a confusion matrix diagram of the Stacking integration analysis framework illustrated according to an exemplary embodiment;

[0022] Figure 6 This is a schematic diagram illustrating a comprehensive index of task processing complexity according to an exemplary embodiment;

[0023] Figure 7 This is a schematic diagram of a process according to an exemplary embodiment;

[0024] Figure 8 This is a principle block diagram illustrated according to an exemplary embodiment;

[0025] Figure 9 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0026] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0027] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Figure 7 An embodiment of a single-task defect identification and analysis method for power equipment according to the present invention is shown.

[0030] In this optional embodiment, the single-task identification and analysis method for power equipment defects includes: Step S101, evaluating the complexity of the single-task data of power equipment defects to be identified by constructing a comprehensive task processing complexity index, and obtaining the task processing complexity; the comprehensive task processing complexity index includes information complexity index, task complexity index, and process complexity index; Step S102, based on the task processing complexity, adaptively allocating differentiated learners by constructing an integrated analysis framework to determine the target identification model corresponding to the single-task data of power equipment defects to be identified; the differentiated learners include random forest model, extreme gradient boosting tree model, and one-dimensional convolutional residual network model; Step S103, according to the task processing complexity and the target identification model, performing collaborative inference scheduling between edge nodes and cloud center, and adaptively allocating the computing power of the target identification model by combining node computing power, network bandwidth, and task execution latency, so as to output the identification and analysis results.

[0031] In an optional embodiment, constructing a comprehensive task processing complexity index includes: using symbol complexity, text complexity, and color complexity as sub-indicators of element complexity, and using layout complexity and element complexity as sub-indicators of information complexity; using main task complexity and task scenario target complexity as sub-indicators of task complexity; using process logic complexity and process step complexity as sub-indicators of process complexity; and generating a comprehensive task processing complexity index based on the information complexity index, task complexity index, and process complexity index; wherein, layout complexity, symbol complexity, text complexity, color complexity, main task complexity, task scenario target complexity, process logic complexity, and process step complexity are all derived by combining entropy theory calculations with expert scoring results.

[0032] In an optional embodiment, the complexity assessment of the single-task data of power equipment defects to be identified includes: establishing an original evaluation index matrix and standardizing it based on the expert scoring results of each complexity index in the comprehensive task processing complexity index to obtain a standardized matrix; calculating the index probability of each complexity index based on the standardized matrix, and calculating the entropy value of each complexity index through the index probability; calculating the entropy redundancy based on the entropy value of each complexity index, and calculating the complexity weight of each complexity index using the entropy redundancy; converting the visual information in the single-task data of power equipment defects to be identified into quantifiable data by constructing a task information structure diagram of power equipment defect identification and analysis, so as to calculate the first-order entropy or second-order entropy corresponding to each complexity index; and using the Euclidean norm to weight the first-order entropy or second-order entropy corresponding to each complexity index to obtain the task processing complexity.

[0033] In an optional embodiment, the expression for task processing complexity is: ; In the formula, H represents the task processing complexity; w j h is the complexity weight of the j-th complexity index; j Let A be the first-order or second-order entropy of the j-th complexity index; i p(A) represents the i-th core information category in the defect identification task. i ) represents the statistical probability of the i-th core information category appearing in historical defect data in the defect identification task.

[0034] In an optional embodiment, constructing an integrated analysis framework includes: determining the corresponding task complexity range based on task processing complexity; generating model allocation instructions based on the task complexity range; determining the target recognition model in the differentiated learner through the model allocation instructions, forming an intelligent fusion and task scheduling layer; using the determined target recognition model, identifying the single-task data of power equipment defects to be identified, obtaining model recognition results that match the task processing complexity, forming a multi-model feature learning layer; and establishing an integrated analysis framework by using the multi-model feature learning layer as the bottom layer and the intelligent fusion and task scheduling layer as the upper layer.

[0035] In an optional embodiment, adaptive allocation of the differential learner includes: when the task processing complexity is in the low complexity range, generating a random forest model allocation instruction and using the random forest model as the target recognition model; when the task processing complexity is in the medium complexity range, or when the single task data of the power equipment defects to be identified has some feature overlap, generating an extreme gradient boosting tree model allocation instruction and using the extreme gradient boosting tree model as the target recognition model; when the task processing complexity is in the high complexity range, generating a one-dimensional convolutional residual network model allocation instruction and using the one-dimensional convolutional residual network model as the target recognition model.

[0036] In an optional embodiment, the random forest model is used as the target recognition model, including: extracting several training subsets from the training sample set and configuring corresponding feature subsets for each training subset; training several decision trees based on the training subsets and feature subsets, and using each decision tree to generate a corresponding defect category judgment result; during the node splitting process of each decision tree, determining the splitting features and splitting thresholds based on the Gini coefficient before and after the node split; and performing majority voting on the defect category judgment results output by different decision trees to obtain the model recognition result corresponding to the random forest model.

[0037] In an optional embodiment, the Gini coefficient is expressed as follows: The expression for obtaining the model identification result corresponding to the random forest model is: In the formula, S is the sample set contained in the node; C is the total number of defect categories; Let H(x) represent the proportion of samples belonging to category c in node S; H(x) is the final prediction result of the random forest model for sample x; h m (x) represents the class prediction result of the m-th decision tree for sample x; I is the indicator function; Y represents the final prediction result of the random forest model for sample x; Y is the set of defect category labels.

[0038] In an optional embodiment, the extreme gradient boosting tree model is used as the target recognition model, including: extracting equipment temperature features, vibration features, and visual features from the single-task data of power equipment defects to be identified, obtaining structured feature data; filtering the structured feature data according to the task processing complexity to obtain the input features, and introducing a class balance coefficient into the input features; the class balance coefficient is generated according to the sample distribution of the defect categories and is used as the sample weight of the training error term in the loss function of the extreme gradient boosting tree model; iteratively correcting the previous round of recognition error under the constraint of the loss function by sequentially training several classification and regression trees; and accumulating the prediction outputs of several classification and regression trees to obtain the model recognition result corresponding to the extreme gradient boosting tree model.

[0039] In an optional embodiment, the expression for obtaining the model identification result corresponding to the extreme gradient boosting tree model is: The expression for the loss function is: In the formula, f is the model prediction value for the i-th sample; t Let x be the t-th classification regression tree; i Let be the i-th input data; K be the number of classification and regression trees; F be the classification and regression tree function space; y i Let f be the true defect category label corresponding to the i-th sample; n is the number of training samples; l is the training error between the predicted value and the true value; Ω is the regularization term; f k Let be the tree function corresponding to the k-th classification regression tree.

[0040] In an optional embodiment, a one-dimensional convolutional residual network model is used as the target recognition model, including: performing convolution processing on the single-task data of power equipment defects to be identified through a one-dimensional convolutional layer to extract local temporal features; inputting the local temporal features into a residual block and combining it with the residual mapping to output a feature map; performing pooling processing on the feature map, and selecting mean pooling or max pooling according to the stability of the feature values ​​within the pooling window; inputting the pooled features into a fully connected classification layer to obtain the category probability corresponding to each defect category; and selecting the defect category with the highest category probability as the model recognition result corresponding to the one-dimensional convolutional residual network model.

[0041] In an optional embodiment, the expression for the output of the residual block is: The expression for class probability is: In the formula, H(X) is the residual block output; X is the input feature; F(X) is the residual mapping; z is the probability that sample x belongs to defect category c; logit,c This represents the linear output value corresponding to defect category c; C is the number of defect categories. This is a temporary variable for summation, used to calculate the sum across all categories.

[0042] In an optional embodiment, collaborative inference scheduling is performed between edge nodes and the cloud center, and the computing power of the target recognition model is adaptively allocated based on node computing power, network bandwidth, and task execution latency. This includes: real-time acquisition of node computing power, network bandwidth, and task execution latency of edge nodes and the cloud center; determining candidate execution paths for inference of the single task data of power equipment defects to be identified on edge nodes or in the cloud center based on task processing complexity and the target recognition model; calculating the inference latency and computing energy consumption corresponding to each candidate execution path to obtain the comprehensive cost of each candidate execution path; selecting the candidate execution path with the minimum comprehensive cost as the target execution path, while ensuring that the inference scheduling process between edge nodes and the cloud center satisfies node load constraints and transmission bandwidth constraints; and calling the target recognition model on the target execution path to perform inference, so as to output the identification and analysis results of the single task of power equipment defects.

[0043] In an optional embodiment, the expression for selecting the candidate execution path with the lowest overall cost as the target execution path is: The expression for the node load constraint is: The expression for the transmission bandwidth constraint is: In the formula, J represents the overall cost; α represents the delay cost weight, and β represents the energy consumption cost weight. In this embodiment, for the scenario of power equipment defect identification, which has high real-time requirements, α = 0.8 and β = 0.2 are taken; T i H represents the task execution delay corresponding to node i; t Let C be the task processing complexity of the t-th defect identification task; i B represents the node computing power of node i; i E represents the network bandwidth corresponding to node i; i Let N be the computational energy consumption corresponding to node i; N be the number of computational nodes participating in collaborative inference scheduling, including edge nodes and cloud centers; M be the number of defect identification tasks within the scheduling period; L be the number of tasks within the scheduling period. it D represents the load of task t on node i; it The bandwidth used for transmitting the t-th defect identification task on node i.

[0044] It should be noted that this invention addresses the problems of complex multimodal data, significant task differences, and limited computing resources in power equipment defect identification, proposing a single-task identification and analysis method for power equipment defects based on multi-model collaboration. For example... Figure 1 As shown, this method focuses on task complexity assessment. It constructs a comprehensive data and task processing complexity index (DTHC) using information entropy and image entropy, quantifying and classifying defect analysis tasks from dimensions such as information complexity, task complexity, and process complexity, providing a basis for model scheduling. Based on this, a Stacking integrated multi-model processing framework integrating random forest, XGBoost, and 1DCNN-ResNet is constructed. This framework adaptively allocates models of different sizes according to tasks of varying complexity, achieving a balance between computational efficiency and recognition accuracy. For application scenarios with high real-time requirements and dynamically changing operating environments, an edge-cloud collaboration mechanism is introduced. Combining factors such as task complexity, node computing power, network bandwidth, and latency, an adaptive computing power allocation and optimization model is established, enabling collaborative scheduling of model inference on the edge and cloud sides. This method effectively reduces computational redundancy, improves defect recognition efficiency and system robustness, and is suitable for intelligent analysis of defects in multi-source, multi-scenario power equipment, thus providing an efficient and scalable solution for power equipment status perception and intelligent operation and maintenance.

[0045] Specifically, this invention provides targeted analysis of the single task of power equipment defect identification based on the differences between multiple models, and the implementation process involves three steps. The first step is to construct a task complexity assessment mechanism, using information entropy and image entropy to assess the complexity of the equipment defect analysis task data and complete task classification. The second step is to construct a task classification processing framework based on Stacking, integrating multiple models, by analyzing the differences in computational performance, storage requirements, and task adaptability between large and small models, and to allocate different models for different tasks. The third step introduces an edge-cloud collaboration framework to dynamically evaluate and schedule task and model resources, optimize the interaction between models and the efficiency of task switching, and combine real-time optimization and dynamic update mechanisms to ensure the system's efficiency and robustness in resource-constrained and dynamically changing scenarios. To achieve the above objectives, this invention adopts the following technical solution:

[0046] 1. Task classification based on complexity assessment:

[0047] The first step is to establish a task handling complexity (DTHC) evaluation mechanism based on information entropy and image entropy according to the complexity of the data source, and to construct a comprehensive evaluation system from the dimensions of information complexity (IC), task complexity (TC) and process complexity (PC).

[0048] 1.1 Establishment of DTHC index:

[0049] Specifically, a comprehensive evaluation index for DTHC is constructed from three dimensions: IC, TC, and PC. IC includes Layout Complexity (ILC) and Element Complexity (EC), with EC further comprising Graphic Complexity (GC), Text Complexity (TEC), and Color Complexity (CC). TC includes Main TC (MTC) and Task Scenario Objective Complexity (TSOC). PC is subdivided into Protocol Logic Complexity (PLC) and Protocol Step Complexity (PSC). The established comprehensive task processing complexity index is as follows: Figure 6 As shown.

[0050] 1.2 Task processing complexity assessment mechanism:

[0051] 1.2.1 Basic Theory of Entropy:

[0052] Specifically, in information theory, entropy is used to describe the degree of disorder or uncertainty of information in a system, representing the average amount of information contained in the system's state. The basic formula for entropy is:

[0053]

[0054] In the formula: H(X) is the entropy of the random variable X; p(x) i ) is event x i The probability of occurrence; log2p(x) i ) represents the event x i The self-information of an image. In the calculation of image entropy, first-order entropy represents the logical structural complexity of an operation, while second-order entropy reflects the information quantity complexity during the operation. Together, they constitute a multi-dimensional description system of information complexity.

[0055] 1.2.2 Complexity Index Weights:

[0056] Specifically, experts with extensive professional knowledge and practical experience are invited to jointly score the performance with the power equipment defect expert system, with scoring levels set from 1 to 5. Assuming there are n experts scoring m complexity indicators, the original evaluation indicator matrix is ​​as follows:

[0057]

[0058] In the formula, x ij Let be the score given by the i-th expert for the j-th complexity index, where i = 1, 2, ..., n; j = 1, 2, ..., m.

[0059] Specifically, the standardized index z ij for:

[0060]

[0061] In the formula: z ij These are the standardized values. The standardized matrix Z = {z ij The entropy value e of the j-th index. j for:

[0062]

[0063] In the formula: p ij For the probability index, ,e j The smaller the value, the greater the dispersion of the j-th indicator, and the more information it contains. Entropy redundancy g j for:

[0064]

[0065] Where: g j This represents the entropy redundancy of the j-th indicator; a larger value indicates a higher importance for that indicator. The weight w of the j-th indicator... j This can be expressed using the following mathematical formula:

[0066]

[0067] 1.2.3 Complexity Assessment Mechanism:

[0068] Specifically, based on the fundamental principles of image entropy, an information structure diagram for power equipment defect identification and analysis tasks with adaptability and accuracy across multiple scenarios can be constructed through information analysis using intelligent algorithms and quantitative evaluation by expert systems. This information structure diagram is a tool for quantifying the complexity of image information. Its core function is to convert the visual information on power equipment defect images into quantifiable data calculated using information entropy. The construction method involves using subsequent intelligent algorithms to analyze the frequency of different defect categories appearing in the power equipment defect images. Then, based on these frequencies, the information entropy is calculated to determine the complexity of the diagram. The first or second entropy h of the j-th index is calculated. j :

[0069]

[0070] In the formula: A i p(A) represents the core information category in the defect identification task. i ) represents information category A i The statistical probability of occurrence in historical defect data.

[0071] Specifically, by weighting each complexity index using the Euclidean norm, the value H of DTHC is obtained as follows:

[0072]

[0073] In the formula: w j This invention represents the complexity weights of the indicators. To improve the scientific rigor and consistency of weight allocation for various complexity indicators in power equipment defect identification and analysis, this invention improves the traditional manual weighting into an expert scoring system that integrates intelligent algorithms. The complexity indicators are derived through entropy theory calculation combined with expert evaluation indicators. This system integrates the hierarchical analysis approach and adaptive optimization algorithms. Through expert scoring data-driven model learning, it automatically generates the final weights of each indicator. Finally, three intelligent algorithm experts in power equipment operation and maintenance are invited to evaluate the results. To ensure the comprehensiveness of the scoring, each indicator is scored from 1 to 5 based on its importance in the power equipment defect identification task to evaluate the importance of each complexity indicator in the DTHC (Difficulty Dimensioning and Computational Hazards). The complexity weights of the eight indicators are then calculated using the above calculation process, as shown in Table 1. Based on the weights in this table and the calculation formula, the DTHC value of a single task can be obtained to complete the complexity classification of the model.

[0074] Table 1 - Weighting of Different Complexity Indicators

[0075]

[0076] 2. Task classification processing method based on stacking ensemble of multiple models:

[0077] 2.1 Stacking Multi-Model Framework:

[0078] It should also be noted that this invention, based on the Stacking ensemble learning framework, adds a model classification processing step based on the task complexity evaluation in the first step, constructing a Stacking ensemble analysis framework oriented towards multi-source features and multi-level task complexity. This method achieves targeted learning and decision optimization for different fault features through the fusion of multiple types of learners and a complexity-driven model allocation mechanism. The overall architecture consists of two layers: a bottom layer is a multi-model feature learning layer, and an upper layer is an intelligent fusion and task scheduling layer.

[0079] To further improve overall recognition efficiency and resource utilization, this invention introduces a first-step task complexity assessment mechanism and a dynamic model allocation mechanism into the framework. When the task complexity is low, the system automatically prioritizes the computationally efficient shallow model, Random Forest, for fast recognition. When the task complexity is moderate or the data features partially overlap, the system schedules the hybrid model, XGBoost, to enhance discriminative ability. For high-complexity tasks, the deep learning module 1DCNN-ResNet is invoked to perform feature extraction and classification. Through this complexity-driven adaptive allocation strategy, the system can significantly reduce computational overhead while maintaining recognition accuracy.

[0080] Based on the aforementioned entropy theory, the task is assigned to the level corresponding to the nearest cluster center by calculating the Euclidean distance between the complexity value H of the task to be classified and the three pre-trained cluster centers.

[0081] Based on the characteristics of power equipment defect identification tasks, the general clustering center is optimized for specific scenarios to obtain a three-level complexity clustering center suitable for power equipment defect analysis:

[0082]

[0083] μ L =0.29 indicates a low-complexity cluster center, which corresponds to a significant deficiency in simple contexts; μ M =0.63

[0084] Cluster centers of medium complexity level correspond to situations where there is some background interference or inconspicuous defect features; μ V =0.85 indicates a high-complexity cluster center, corresponding to minor defects or multiple defects superimposed in a complex background.

[0085] Calculate the absolute distance d between the complexity H of the classification task and each cluster center. k The rank corresponding to the cluster center with the smallest distance is selected as the final classification result:

[0086]

[0087] In the formula, μ k It is the one-dimensional cluster center value of the kth complexity level, L ∗ This represents the complexity level of the output, with values ​​L, M, and N, corresponding to low, medium, and high complexity, respectively.

[0088] In the second step, the feature learning stage, this invention introduces multiple differentiated learners to comprehensively extract features from the operating signals of power equipment. A deep learning model is used to capture nonlinear local features in the original waveforms, images, or time-series signals. The deep learning part employs an improved convolutional residual network structure, maintaining gradient stability and enhancing deep feature abstraction capabilities through residual mapping, making it suitable for signal recognition in complex electromagnetic noise environments. The machine learning part enhances the model's discriminative ability through error weighting and feature randomization. After parallel training of multiple learners, their output is used as input for the third step, the edge-cloud collaborative optimization part.

[0089] 2.2 Multi-model algorithms for different levels of complexity:

[0090] 2.2.1 Random Forest Algorithm:

[0091] Specifically, the Random Forest algorithm is an ensemble classification algorithm based on the Bagging method, which improves accuracy by combining the classification results of multiple decision trees. Each decision tree is trained using randomly selected datasets and features, and the optimal classification result is determined through a voting mechanism. In power defect identification, this invention employs this algorithm to handle low-complexity tasks because it does not require tuning too many parameters and is suitable for processing multimodal data.

[0092] The original dataset used in this invention is the Kaggle open-source fault dataset, with a total of 8000 samples. The defect types include insulator string failure, transformer corrosion and leakage, equipment wire breakage, and casing cracks. The dataset is divided into training and testing sets in an 8:2 ratio.

[0093] Specifically, the Random Forest algorithm first extracts M independent subsets from the original dataset to provide differentiated training data for each decision tree. Each tree is independently trained based on its own training subset and feature subset, generating a power defect type judgment for the current input sample. The feature subset ensures that each decision tree identifies equipment defects differently, improving the model's generalization ability. In the final prediction stage, each tree votes on the defect category of the input sample. After summing the voting results of all trees, the category with the most votes becomes the defect identification result output by the system. In the power defect identification scenario, a single decision tree may not be stable enough in identifying certain defects (such as tiny cracks in a complex background), but the ensemble of multiple trees can significantly improve the overall accuracy and generalization ability of the identification. For the m-th decision tree, let it be defined as D... m Each tree is trained using only a subset of features to reduce inter-tree correlation. Features are randomly selected from all D features of the decision tree. Each feature forms a feature subset F. mThe decision tree then improves purity by recursively splitting nodes, using the Gini coefficient to measure node purity, and selecting the feature and threshold that result in the highest purity after splitting as the split point.

[0094] Specifically, let a node contain a sample set of... Then the Gini coefficient of node S is defined as:

[0095]

[0096] In the formula: The percentage of samples belonging to category c in node S; Let S be the total number of samples contained in node S. In the process of constructing a decision tree for identifying defects in power equipment, the first step is to analyze the various equipment state samples contained in the current tree node, such as image and signal feature data of insulators and transformers. The Gini coefficient is calculated to measure the proportion of samples of different equipment defect types and normal states within that node. If all samples within the node are of the same defect type, the Gini value is the lowest, indicating the highest purity. If multiple types of samples are included, such as cracks, corrosion, and normal states, the purity is poor. During splitting, the algorithm sequentially tests each defect-sensitive feature from a randomly selected feature subset and divides each feature into left and right sub-features. By comparing the weighted Gini values ​​of the sub-nodes after each split, the feature and threshold that most clearly classify the defect types within the left and right sub-nodes are selected as the split point, thus allowing the decision tree to accurately distinguish different equipment defects.

[0097] Specifically, for node S, traverse the feature subset F m Each feature in And try all possible thresholds t for this feature, splitting node S into left child node S. L and right child node S R The process recursively splits nodes until all node samples belong to the same category or the tree depth reaches a preset maximum value, at which point the process terminates. After training, the random forest merges the prediction results of the M trees through majority voting, reducing the risk of overfitting from a single tree and improving the model's generalization ability.

[0098] Specifically, for any sample x∈ RD The final prediction function H(x) of a random forest is defined as:

[0099]

[0100] In the formula: Let I() represent the class prediction of the m-th tree for sample x. If I = 1, then I = 1; otherwise I = 0. This represents the number of trees that received votes for category c. Ultimately, the category with the most votes is selected as the result of the power equipment defect detection.

[0101] 2.2.2 XGBoost Algorithm:

[0102] Specifically, for low-to-medium difficulty tasks with low complexity and small data volume, this invention employs the Extreme Gradient Boosting Tree (XGBoost) algorithm. XGBoost is an improvement on the GBDT algorithm. It constructs a temperature feature column by extracting statistical quantities such as maximum, minimum, mean, and standard deviation of temperature; it extracts frequency domain features from vibration time-domain signals, including dominant frequency amplitude and harmonic proportion, to construct a vibration feature column; and it uses edge detection and color moment methods to extract fault area, aspect ratio, and mean color histogram values ​​from inspection images to construct a visual feature column. This is suitable for processing structured feature data such as equipment temperature, vibration frequency, and image extraction from power inspection data.

[0103] Compared to the traditional XGBoost algorithm, this invention combines power equipment defect features with the aforementioned complexity assessment method, making the following improvements: First, dynamic feature filtering filters input features based on the derived task processing complexity index, reducing computational overhead. Second, a class balance coefficient is introduced. The class balance coefficient is used to improve the performance of less frequent but more critical defect types, addressing the class imbalance problem in power defect samples. GBDT uses greedy learning, fitting the residuals with a CART tree in each iteration. XGBoost, on the other hand, differs by performing a second-order Taylor expansion of the function and adding a regularization term to reduce overfitting risk, and utilizing multi-threaded parallel computation to significantly improve model training speed. Its algorithm model is as follows:

[0104]

[0105] In the formula: f is the model prediction value for the i-th sample; K is the number of trees; t This represents the structure corresponding to the t-th independent tree; x i f is the i-th input data; t (x i ) represents the predicted output of the t-th tree to the input; F represents the set of all possible CARTs.

[0106] Specifically, the XGBoost algorithm optimizes the model using an incremental training strategy, further reducing the loss function size until all K trees have been optimized. In power grid defect identification, input data may include fault feature images extracted from captured images or a small set of fault parameters. The mathematical formula for the XGBoost loss function is:

[0107]

[0108] In the formula: The training error between the predicted value and the target true value. This is a regularization term used to prevent overfitting of the model, especially effective when power image data is noisy. Class balance coefficient. The determination is dynamically based on the frequency of occurrence of the defect type of the sample in the training set:

[0109]

[0110] Where: N total C is the total number of training samples; C is the total number of defect categories; N is the total number of training samples. ci Let c be the category to which sample i belongs. i The number of samples.

[0111] Regular terms The mathematical formula is as follows:

[0112]

[0113] In the formula: γ is used to control the number of leaf nodes to prevent the CART tree from becoming too deep and sensitive to noise; λ is a control coefficient to avoid extreme predictions; ω j denoted by , where is the score of the leaf nodes; and T is the number of leaf nodes.

[0114] It should be added that XGBoost is suitable for tasks of low to medium complexity, with a time complexity of O(n log n). n K), where n is the number of samples and K is the number of trees. In low-to-medium complexity power defect identification, directly allocating XGBoost can save computational resources.

[0115] 2.2.3, 1DCNN-ResNet deep learning model:

[0116] Specifically, the one-dimensional convolutional residual network model used in this invention, namely 1DCNN-ResNet, is a combination of one-dimensional convolutional neural network and residual network. It can be used for high-complexity tasks with high accuracy requirements and large data volume. Its main structure consists of 1D convolutional layer, residual block, pooling layer, and fully connected classification layer.

[0117] 2.2.3.1 1D Convolutional Layer:

[0118] Specifically, the 1D convolutional layer extracts local temporal features by sliding a one-dimensional convolutional kernel through the input data, serving as the basic feature extraction module. For the input feature map... Output feature map of 1D convolutional layer The calculation is as follows:

[0119]

[0120] In the formula: C in Where Cout is the number of channels in the input feature map, and W is the number of output convolution kernels. d,c,q b represents the weight parameters of the 1D convolution kernel. q Let L be the bias parameter of the q-th convolutional kernel, ReLU be the rectified linear unit activation function, and L be the length of the output feature map. out The input length L, kernel size k, stride s, and padding p are used to determine the input length L, kernel size k, stride s, and padding p. pad Decide:

[0121]

[0122] 2.2.3.2, Residual Block:

[0123] Specifically, the core of ResNet is the residual block. The signal features of some defects in the power system need to be effectively extracted through deep networks. However, when the depth of traditional 1DCNN increases, gradient vanishing occurs. The residual block solves this problem through identity mapping.

[0124] Specifically, the output H(X) of the residual block is the sum of the residual mapping F(X) and the input X:

[0125]

[0126]

[0127] In the formula: F(X) contains two 1D convolutional layers, a BN layer, and ReLU activation; W1 and W2 are the weights of the two 1D convolutional layers in the residual mapping; b1 and b2 are the biases of the two convolutional layers; and BN is the batch normalization function. To address the characteristics of blurred edges and strong background interference in power equipment image defect samples, the convolutional kernel size is reduced in the residual mapping branch to preserve information about broken wire ends, crack tips, and insulator edges. The residual block can stably capture the discriminative visual features of small-sized defects in power equipment images while increasing network depth.

[0128] 2.2.3.3 Pooling Layer:

[0129] Specifically, the pooling layer addresses the time-series offset and signal redundancy issues in defect signals from power equipment through downsampling, reducing computational load, adapting to the real-time analysis needs of power systems, and enhancing the model's robustness to time-series offsets. Instead of simply selecting local maxima, the pooling operation first calculates the variance of the signal amplitude within each pooling window. If the variance is below a preset threshold, the window is considered to be in a stationary redundancy state. In this case, the background vibration or steady-state temperature signal from normal power equipment operation is replaced with the window mean as the pooling output to suppress noise disturbances. If the variance is high, it indicates that the window contains effective transient characteristic signals caused by defects such as partial discharge, mechanical structural abnormalities, or insulation breakdown. In this case, max pooling is retained to highlight abnormal peak values ​​caused by defects. The calculation formula is as follows:

[0130]

[0131] For a one-dimensional time series value x within the window 1 arrive x n First calculate the window mean μ, then calculate the variance σ, where n is the window length.

[0132] It should also be noted that this processing method effectively compresses a large amount of redundant monitoring data under steady-state conditions, reduces the size of feature maps and subsequent computational overhead, and makes the model more robust to time-series shifts caused by electromagnetic interference or operating condition fluctuations. The preset threshold is determined through statistical analysis of historical defect data. Monitoring signal samples from various devices are collected during normal steady-state operation, and the distribution of their pooling window variance is calculated. The upper bound of the lower difference quantile under normal operating conditions is taken as the initial threshold, ensuring that the variance of the vast majority of stable windows is below this value.

[0133] Specifically, for the input feature map Output of one-dimensional data max pooling for:

[0134]

[0135] In the formula: For the input feature map at location The eigenvalues ​​on channel q.

[0136] 2.2.3.4 Fully Connected Classification Layer:

[0137] Specifically, the role of the fully connected classification layer is to... High-dimensional features are transformed into distinct defect types, that is, the high-dimensional features output by the pooling layer are mapped into category probability vectors to classify power defect types. The fully connected classification layer first flattens the multi-dimensional feature map after pooling into a one-dimensional vector, which summarizes the waveform features of defects such as partial discharge or overheating extracted from the signal. Then, it is fully connected and mapped to the same dimension as the number of defect categories. Each row of the weight matrix corresponds to the response template of a specific defect type, such as insulator contamination or equipment casing cracks. The similarity score between the input waveform and each type of template is obtained through dot product, and then normalized to a probability distribution using Softmax. The category with the highest probability is taken as the recognition result, thus accurately distinguishing different defects. The mathematical model is as follows: first, a linear transformation is performed in the fully connected layer, flattening the feature map Z output by the pooling layer into a vector z, and then performing a linear transformation again.

[0138]

[0139] Then, softmax activation is performed to obtain the class probability output, and the linear output z is then processed. logit Mapped to a probability distribution:

[0140]

[0141] Finally, perform a final category prediction and select the category with the highest probability. As a defect type:

[0142]

[0143] In the formula: W fc Let b be the weight matrix of the fully connected layer. fc z is the bias vector of the fully connected layer. logit For the linear output of the fully connected layer, z logit,c Let P(c|x) be the linear output value of the c-th defect, and let P(c|x) be the probability that sample x belongs to the c-th defect.

[0144] 3. Edge-cloud collaboration optimization:

[0145] It should be noted that, based on the task complexity assessment and multi-model allocation mechanism in the first two chapters, this invention proposes a multi-model optimization and resource scheduling technology framework based on edge-cloud collaboration, which optimizes the interaction between models and the efficiency of task switching. Combined with real-time optimization and dynamic update mechanisms, it ensures the efficiency and robustness of the system in resource-constrained and dynamically changing scenarios.

[0146] In the intelligent identification of the operating status of power equipment, different tasks have significantly different requirements for model complexity and computing resources. To address this characteristic, this invention employs a multi-model hierarchical inference mechanism, constructing a collaborative inference system between edge nodes and the cloud center. When the system detects a low task complexity, edge nodes prioritize calling random forest or lightweight XGBoost models for rapid identification and directly feed back the results. For high-complexity tasks, the system directly offloads the raw data or deep features to the cloud, where 1DCNN-ResNet performs comprehensive inference analysis.

[0147] To ensure the efficient operation of the edge-cloud collaborative system in multi-task parallel and dynamic scenarios, this invention introduces an adaptive computing power allocation model based on multi-factor feedback. The system monitors the task complexity index H in real time. t Node computing power capacity C i Network bandwidth B i And the task execution delay T of node i. i A computing power allocation optimization model is constructed. Specifically, when the system receives a power inspection task, it collects four key parameters in real time: the task complexity index based on data calculation, the available computing power capacity C of the edge computing nodes, and so on. i The current network bandwidth of the backhaul channel, B i And the historical task execution delay T for identifying similar defects i .

[0148] Specifically, each time the task is assigned, the total inference latency T corresponding to assigning the task to the edge or the cloud is calculated. total With the total energy consumption E total Based on the system's requirements for speed and energy efficiency, the execution path with the lowest overall cost is selected. Through real-time detection and correction, the system continuously optimizes the overall inference latency and total computational energy consumption. The overall objective function is to minimize the practical T. total With energy consumption E total The goal is to minimize the overall system inference latency and energy consumption.

[0149]

[0150] In the formula: α and β are weighting coefficients, which reflect the system's preference for real-time performance and energy consumption, respectively.

[0151] Specifically, the constraints of edge-cloud collaboration include node load constraints and transmission bandwidth constraints; among them, the node load constraint is the node's task load and available computing power C. i Must meet:

[0152]

[0153] The network's transmission bandwidth constraint is:

[0154]

[0155] In the formula: L it This represents the load of task t on node i. Where D it This refers to the bandwidth used for transmitting task data between nodes.

[0156] Furthermore, through the aforementioned optimization model and iterative mechanism, this invention can achieve adaptive allocation and real-time optimization of computing power under different network conditions and task loads, thereby reducing inference latency, improving identification efficiency, and ensuring the continuity and stability of the power equipment defect analysis process.

[0157] To facilitate understanding of the above technical solution of the present invention, the following detailed explanation uses a power equipment defect identification dataset as an example:

[0158] In this power equipment defect identification dataset, the power equipment defect samples to be identified are divided according to the true defect categories, and identification tests are conducted using the 1DCNN-ResNet algorithm, XGBoost algorithm, Random Forest algorithm, and the Stacking ensemble analysis framework corresponding to this invention. The diagnostic results of each model are then displayed in the form of a confusion matrix. In the confusion matrix, the horizontal axis represents the predicted category, and the vertical axis represents the true category. The values ​​in the matrix represent the number of true category samples that are predicted as the corresponding category, thus intuitively reflecting the accuracy of different models in identifying various types of power equipment defects.

[0159] like Figures 2 to 5 As shown, Figure 2 The confusion matrix of the 1DCNN-ResNet algorithm shows that the model performs well in class 2, with all 100 class 2 samples being correctly identified. However, 15 samples in class 1 are misclassified as class 0, and 10 samples in class 0 are misclassified as class 1. Figure 3 The image shows the confusion matrix of the XGBoost algorithm. It can be seen that this model can also accurately identify category 2, and the number of misclassifications for category 1 and category 0 is significantly lower. Figure 2 The number of cases decreased, with 11 samples from category 1 being misclassified as category 0 and 9 samples from category 0 being misclassified as category 1. Figure 4 The confusion matrix of the random forest algorithm shows that the model further improves its ability to distinguish between category 0 and category 1. 10 samples in category 1 were misclassified as category 0, and 8 samples in category 0 were misclassified as category 1. Figure 5The image shows the confusion matrix of the Stacking ensemble analysis framework in this invention. It can be seen that this invention further reduces the number of misclassifications in categories 0 and 1. Only 7 samples in category 1 were misclassified as category 0, and only 4 samples in category 0 were misclassified as category 1. This indicates that this invention can significantly improve the differentiation effect between different defect categories through multi-model collaboration.

[0160] Meanwhile, to further compare the running efficiency of different models, the running time of each model on the same defect identification dataset was statistically analyzed, and the results are shown in Table 2.

[0161] Table 2 - Comparison of computation time for defect identification and classification models

[0162]

[0163] The runtime of the 1DCNN-ResNet algorithm is 0.9957s, the XGBoost algorithm is 1.0586s, the Random Forest algorithm is 1.7145s, and the Stacking ensemble model is 1.1485s. As shown in Table 2, although the Stacking ensemble analysis framework of this invention integrates multiple differential learners, its runtime is still within an acceptable range, and it has a shorter runtime than the Random Forest algorithm, thus ensuring both recognition performance and inference efficiency.

[0164] Figure 8 An embodiment of a single-task defect identification and analysis system for power equipment according to the present invention is shown.

[0165] In this optional embodiment, the power equipment defect single-task identification and analysis system includes: a complexity assessment module 201, used to assess the complexity of the power equipment defect single-task data to be identified by constructing a comprehensive task processing complexity index, thereby obtaining the task processing complexity; the comprehensive task processing complexity index includes information complexity index, task complexity index, and process complexity index; a model allocation module 202, used to adaptively allocate differentiated learners based on the task processing complexity by constructing an integrated analysis framework, thereby determining the target identification model corresponding to the power equipment defect single-task data to be identified; the differentiated learners include a random forest model, an extreme gradient boosting tree model, and a one-dimensional convolutional residual network model; and a collaborative scheduling module 203, used to perform collaborative inference scheduling between edge nodes and the cloud center according to the task processing complexity and the target identification model, and adaptively allocate the computing power of the target identification model by combining node computing power, network bandwidth, and task execution latency, so as to output the identification and analysis results.

[0166] In an optional embodiment, the model allocation module 202, when constructing the integrated analysis framework, includes: determining the corresponding task complexity range based on the task processing complexity; generating a model allocation instruction based on the task complexity range; determining the target recognition model in the differentiated learner through the model allocation instruction, forming an intelligent fusion and task scheduling layer; using the determined target recognition model, identifying the single task data of the power equipment defects to be identified, obtaining a target recognition model that matches the task processing complexity, forming a multi-model feature learning layer; and establishing an integrated analysis framework by using the multi-model feature learning layer as the bottom layer and the intelligent fusion and task scheduling layer as the upper layer.

[0167] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0168] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0169] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0170] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0172] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A single-task defect identification and analysis method for power equipment, characterized in that, The single-task defect identification and analysis method for power equipment includes: By constructing a comprehensive task processing complexity index, the complexity of the single task data of the power equipment defects to be identified is evaluated to obtain the task processing complexity; the comprehensive task processing complexity index includes information complexity index, task complexity index and process complexity index. Based on the task processing complexity, an integrated analysis framework is constructed to adaptively allocate differential learners and determine the target recognition model corresponding to the single task data of the power equipment defects to be identified. The differential learners include a random forest model, an extreme gradient boosting tree model, and a one-dimensional convolutional residual network model. Based on the task processing complexity and the target recognition model, collaborative reasoning scheduling is performed between edge nodes and the cloud center. The computing power of the target recognition model is adaptively allocated by combining node computing power, network bandwidth and task execution latency to output recognition analysis results.

2. The method for single-task identification and analysis of defects in power equipment according to claim 1, characterized in that, The comprehensive index of the complexity of the construction task includes: Symbol complexity, text complexity, and color complexity are used as sub-indices of element complexity, and layout complexity and element complexity are used as sub-indices of the information complexity index. The complexity of the main task and the complexity of the task scenario objectives are used as subordinate indicators of the task complexity index. The complexity of process logic and the complexity of process steps are used as sub-indicators of the process complexity index. Based on the information complexity index, task complexity index, and process complexity index, a comprehensive task processing complexity index is generated. The complexity of the layout, the complexity of the symbols, the complexity of the text, the complexity of the colors, the complexity of the main task, the complexity of the task scenario objectives, the complexity of the process logic, and the complexity of the process steps are all derived by combining entropy theory calculations with expert scoring results.

3. The single-task defect identification and analysis method for power equipment according to claim 1, characterized in that, The complexity assessment of the single-task data for identifying power equipment defects includes: Based on the expert scores of each complexity index in the comprehensive task processing complexity index, an original evaluation index matrix is ​​established and standardized to obtain a standardized matrix. The probability of each complexity index is calculated based on the standardized matrix, and the entropy value of each complexity index is calculated based on the index probability. Entropy redundancy is calculated based on the entropy values ​​of various complexity indicators, and the complexity weight of each complexity indicator is calculated using the entropy redundancy. By constructing a task information structure diagram for power equipment defect identification and analysis, the visual information in the single task data of the power equipment defects to be identified is converted into quantifiable data in order to calculate the first-order entropy or second-order entropy corresponding to each complexity index. The task processing complexity is obtained by weighting the first-order or second-order entropy corresponding to each complexity index using the Euclidean norm.

4. The single-task defect identification and analysis method for power equipment according to claim 3, characterized in that, The expression for the task processing complexity is: ； ； In the formula, H represents the task processing complexity; w j h is the complexity weight of the j-th complexity metric; j Let A be the first-order or second-order entropy of the j-th complexity index; i p(A) represents the i-th core information category in the defect identification task. i ) represents the statistical probability of the i-th core information category appearing in historical defect data in the defect identification task.

5. The method for single-task identification and analysis of defects in power equipment according to claim 1, characterized in that, The construction of the integrated analysis framework includes: The task complexity range is determined by the task processing complexity, a model allocation instruction is generated based on the task complexity range, and a target recognition model is determined in the differential learner by the model allocation instruction, thus forming an intelligent fusion and task scheduling layer. Using a defined target recognition model, the single-task data of the power equipment defects to be identified is used to obtain model recognition results that match the processing complexity of the task, thus forming a multi-model feature learning layer. An integrated analysis framework is established by using a multi-model feature learning layer as the bottom layer and an intelligent fusion and task scheduling layer as the top layer.

6. The method for single-task identification and analysis of defects in power equipment according to claim 5, characterized in that, The adaptive allocation of differential learners includes: When the task processing complexity is in the low complexity range, a random forest model allocation instruction is generated, and the random forest model is used as the target recognition model. When the task processing complexity is in the medium complexity range, or when the single task data of the power equipment defects to be identified have some overlapping features, an extreme gradient boosting tree model allocation instruction is generated, and the extreme gradient boosting tree model is used as the target identification model. When the task processing complexity is in the high complexity range, a one-dimensional convolutional residual network model allocation instruction is generated, and the one-dimensional convolutional residual network model is used as the target recognition model.

7. The method for single-task identification and analysis of defects in power equipment according to claim 6, characterized in that, The use of a random forest model as a target recognition model includes: Several training subsets are extracted from the training sample set, and a corresponding feature subset is configured for each training subset; Several decision trees are trained based on the training subset and feature subset, and each decision tree is used to generate a corresponding defect category judgment result; During the node splitting process of each decision tree, the splitting characteristics and splitting threshold are determined based on the Gini coefficient before and after the node splitting. The defect category judgment results output by different decision trees are subjected to majority voting to obtain the model recognition result corresponding to the random forest model.

8. The method for single-task identification and analysis of defects in power equipment according to claim 7, characterized in that, The expression for the Gini coefficient is as follows: ； The expression for obtaining the model identification result corresponding to the random forest model is: ； In the formula, S is the sample set contained in the node; C is the total number of defect categories; Let H(x) represent the proportion of samples belonging to category c in node S; H(x) is the final prediction result of the random forest model for sample x; h m (x) represents the class prediction result of the m-th decision tree for sample x; I is the indicator function; Y represents the final prediction result of the random forest model for sample x; Y is the set of defect category labels.

9. The single-task defect identification and analysis method for power equipment according to claim 6, characterized in that, The method of using an extreme gradient boosting tree model as a target recognition model includes: The equipment temperature features, vibration features, and visual features are extracted from the single-task data of the power equipment defects to be identified to obtain structured feature data; The structured feature data is filtered according to the task processing complexity to obtain the input features, and a class balance coefficient is introduced into the input features. The class balance coefficient is generated according to the sample distribution of the defect category and is used as the sample weight of the training error term to be introduced into the loss function of the extreme gradient boosting tree model. By training several classification and regression trees in sequence, the recognition error of the previous round is iteratively corrected under the constraint of the loss function; The predicted outputs of several classification and regression trees are summed to obtain the model recognition result corresponding to the extreme gradient boosting tree model.

10. The single-task defect identification and analysis method for power equipment according to claim 9, characterized in that, The expression for obtaining the model identification result corresponding to the extreme gradient boosting tree model is: ； The expression for the loss function is: ； In the formula, f is the model prediction value for the i-th sample; t Let x be the t-th classification regression tree; i Let be the i-th input data; K be the number of classification and regression trees; F be the function space of the classification and regression trees; y i Let f be the true defect category label corresponding to the i-th sample; n is the number of training samples; l is the training error between the predicted value and the true value; Ω is the regularization term; f k Let be the tree function corresponding to the k-th classification regression tree.

11. The single-task defect identification and analysis method for power equipment according to claim 6, characterized in that, The method of using a one-dimensional convolutional residual network model as a target recognition model includes: The single-task data of the power equipment defects to be identified is processed by a one-dimensional convolutional layer to extract local temporal features. Local temporal features are input into the residual block, and the feature map is output by combining the residual mapping. The feature map is pooled, and mean pooling or max pooling is selected based on the stability of the feature values ​​within the pooling window. The pooled features are input into a fully connected classification layer to obtain the class probability corresponding to each defect category. The defect category with the highest probability is selected as the model recognition result corresponding to the one-dimensional convolutional residual network model.

12. The single-task defect identification and analysis method for power equipment according to claim 11, characterized in that, The expression for the output of the residual block is: ； The expression for the category probability is: ； In the formula, H(X) is the residual block output; X is the input feature; F(X) is the residual mapping; z is the probability that sample x belongs to defect category c; logit,c This represents the linear output value corresponding to defect category c; C is the number of defect categories. This is a temporary variable for summation, used to calculate the sum across all categories.

13. The single-task defect identification and analysis method for power equipment according to claim 1, characterized in that, The method of collaborative inference scheduling between edge nodes and the cloud center, and adaptively allocating the computing power of the target recognition model based on node computing power, network bandwidth, and task execution latency, includes: Real-time acquisition of node computing power, network bandwidth, and task execution latency of edge nodes and cloud centers; Based on the task processing complexity and target recognition model, candidate execution paths for inference of the single task data of the power equipment defects to be identified are determined at the edge node or cloud center. Calculate the inference latency and computational energy consumption for each candidate execution path to obtain the comprehensive cost of each candidate execution path; The candidate execution path with the lowest overall cost is selected as the target execution path, while ensuring that the inference scheduling process between edge nodes and the cloud center meets the node load constraints and transmission bandwidth constraints. The target recognition model is invoked on the target execution path to perform inference, so as to output the identification and analysis results of the power equipment defect single task.

14. The single-task defect identification and analysis method for power equipment according to claim 13, characterized in that, The expression for selecting the candidate execution path with the lowest overall cost as the target execution path is: ； The expression for the node load constraint is: ； The expression for the transmission bandwidth constraint is: ； In the formula, J represents the overall cost; α represents the delay cost weight; β represents the energy consumption cost weight; T i H represents the task execution delay corresponding to node i; t Let C be the task processing complexity of the t-th defect identification task; i B represents the node computing power of node i; i E represents the network bandwidth corresponding to node i; i Let N be the computational energy consumption corresponding to node i; N be the number of computing nodes participating in collaborative inference scheduling, including edge nodes and cloud centers; M be the number of defect identification tasks within the scheduling period; L be the number of tasks within the scheduling period. it D represents the load of task t on node i; it The bandwidth used for transmitting the t-th defect identification task on node i.

15. A single-task defect identification and analysis system for power equipment, characterized in that, This power equipment defect single-task identification and analysis system includes: The complexity assessment module is used to assess the complexity of the single task data of the power equipment defects to be identified by constructing a comprehensive task processing complexity index, thereby obtaining the task processing complexity; the comprehensive task processing complexity index includes information complexity index, task complexity index and process complexity index. The model allocation module is used to adaptively allocate differential learners based on the task processing complexity by constructing an integrated analysis framework, and determine the target recognition model corresponding to the single task data of power equipment defects to be identified; the differential learners include random forest model, extreme gradient boosting tree model and one-dimensional convolutional residual network model. The collaborative scheduling module is used to perform collaborative reasoning scheduling between edge nodes and the cloud center based on the task processing complexity and the target recognition model, and adaptively allocate the computing power of the target recognition model in combination with node computing power, network bandwidth and task execution latency, so as to output the recognition analysis results.

16. A single-task defect identification and analysis system for power equipment according to claim 15, characterized in that, The model allocation module, when constructing the integrated analysis framework, includes: The task complexity range is determined by the task processing complexity, a model allocation instruction is generated based on the task complexity range, and a target recognition model is determined in the differential learner by the model allocation instruction, thus forming an intelligent fusion and task scheduling layer. Using a defined target recognition model, the single-task data of the power equipment defects to be identified is used to obtain a target recognition model that matches the processing complexity of the task, thus forming a multi-model feature learning layer. An integrated analysis framework is established by using a multi-model feature learning layer as the bottom layer and an intelligent fusion and task scheduling layer as the top layer.

Images