Data-knowledge fusion driven power equipment state evaluation method and system
Through the data-knowledge fusion-driven method, Raindrop learning algorithm and graph neural network are used to process irregular feature parameter data, and a health assessment model is built, which solves the problems of data irregularity and insufficient samples in the health status evaluation of power equipment, achieving high efficiency Accurate health status assessment.
Patent Information
- Application Number
- CN202410904207.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2024-11-12
AI Technical Summary
The existing methods of power equipment health status assessment rely on manual inspection and online monitoring, resulting in irregular characteristic parameter data, errors in measurement or loss of transmission, reducing the accuracy of evaluation, and fewer precise labeling samples, making it difficult to achieve efficient and accurate health. Status Assessment.
Using a data-knowledge fusion-driven method, the Raindrop learning algorithm is used to build a health assessment model based on strong annotation samples and knowledge experience, and to process irregular feature parameter data through graph neural network and attention mechanism, build a directed weighted graph, and update edge weights. Introduce knowledge regularization terms and knowledge functions to enhance the generalization ability of the model.
Improve the accuracy and robustness of the health assessment of power equipment, better adapt to the health assessment needs of actual power equipment, overcome the problems of irregularity of characteristic parameter data and insufficient samples, and improve the generalization ability of the model and Computational efficiency.
Smart Images

Figure CN118940143A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power fault diagnosis and intelligent operation and maintenance, and in particular to a method and system for evaluating the state of power equipment driven by data-knowledge fusion. Background Art
[0002] At present, the health status assessment of power equipment is an important prerequisite for improving reliability and one of the basic tasks for achieving lean operation and maintenance control. With the rapid development of photovoltaic / wind power generation, electric vehicles, and energy storage, the operation of power equipment is facing the superposition of new source and load power impacts, deterioration of power quality, and complex environmental conditions. The health status of power equipment is accelerating, and the number of defect faults such as insulation performance degradation, enhanced partial discharge, oil temperature, and joint overheating has increased significantly. With the current increase in the number of power equipment year by year, the overall availability coefficient is on a downward trend. In addition, the construction and development of the power Internet of Things in recent years has laid a data foundation for obtaining the status parameters of power equipment. In order to improve the reliability of power equipment and the benefits of lean management throughout the life cycle, it is necessary to break the traditional distribution operation and maintenance control mode such as manual inspection and post-event repair, and build a new operation and maintenance system with prevention as the main focus, forward-moving checkpoints, active and efficient, and system control. There is an urgent need to efficiently and accurately assess the health status of power equipment to promote the transformation of power equipment operation and maintenance from traditional passive repair after failure to active operation and maintenance control.
[0003] The premise for the existing methods to achieve accurate assessment of the health status of power equipment is to obtain a complete and effective set of characteristic parameters. However, in practice, different characteristic parameters are often obtained through different methods such as manual inspection and online monitoring. The sampling time interval of manual inspection is not fixed, and there are problems such as characteristic parameter measurement errors or transmission loss in online monitoring. The actual data is irregular, which significantly reduces the accuracy of existing health assessment methods. Summary of the invention
[0004] In order to solve the problem that the characteristic parameter set required for the existing power equipment health status assessment is mainly obtained by sampling through manual inspection or online monitoring, the sampling time interval of manual inspection is not fixed, and there are problems such as characteristic parameter measurement errors or transmission loss in online monitoring, and the actual data is irregular and there are few accurately labeled samples, which leads to the reduction of the accuracy of the power equipment health assessment method, the present invention proposes a data-knowledge fusion driven power equipment status assessment method, including:
[0005] Acquire characteristic parameter data of power equipment;
[0006] Based on the characteristic parameter data of the power equipment, a pre-built health assessment model of the power equipment is solved using a Raindrop learning algorithm to obtain a health status level of the power equipment;
[0007] The health assessment model is constructed based on strongly labeled samples of the power equipment acquired in advance and weakly labeled samples obtained by formally expressing the knowledge and experience of the power equipment.
[0008] Optionally, the health assessment model of the power equipment includes the following construction process:
[0009] Obtain strong annotation samples and knowledge and experience of the power equipment;
[0010] Formalize the knowledge and experience to obtain weakly labeled samples;
[0011] According to the strongly labeled samples, construct a cross entropy loss function of the strongly labeled samples;
[0012] According to the weakly labeled samples, construct a knowledge regularization term of the weakly labeled samples;
[0013] Constructing a knowledge function according to the strongly labeled samples and the weakly labeled samples;
[0014] Obtaining a cross entropy loss function of the health assessment model of the power equipment according to the cross entropy loss function of the strongly labeled samples, the knowledge regularization term of the weakly labeled samples and the knowledge function;
[0015] The health assessment model of the power equipment is obtained with the goal of minimizing the cross entropy loss function of the health assessment model of the power equipment.
[0016] Optionally, constructing a knowledge function according to the strongly labeled samples and the weakly labeled samples includes:
[0017] According to the strongly labeled samples and the weakly labeled samples, the strength of the characteristic relationship between the characteristic parameter data in the electric power equipment is obtained;
[0018] According to the strength of the characteristic relationship between the characteristic parameter data in the power equipment, a knowledge function corresponding to each set of characteristic relationships is constructed.
[0019] Optionally, the expression corresponding to the cross entropy loss function of the health assessment model is as follows:
[0020]
[0021] Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i)represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ 1 represents a hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 Indicates the weight coefficient value corresponding to the first group of feature relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationship between characteristic parameter data α and characteristic parameter data β; l=1…n; n represents the number of groups of characteristic parameter data forming the characteristic relationship.
[0022] Optionally, the knowledge function E corresponding to the first set of characteristic relations between the characteristic parameter data α and the characteristic parameter data β is l The expression of (α,β) is as follows:
[0023]
[0024] Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,n}; n represents the number of groups of feature parameter data forming the feature relations; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data at time X; Ω G represents the sampling set of characteristic parameter data at time G; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value.
[0025] Optionally, after obtaining the strength of the characteristic relationship between each characteristic parameter data in the power equipment according to the strongly labeled samples and the weakly labeled samples, before constructing the knowledge function corresponding to each group of characteristic relationships according to the strength of the characteristic relationship between each characteristic parameter data in the power equipment, the method further includes:
[0026] Taking each of the feature parameter data as a vertex, taking the feature relationship between each of the feature parameter data as an edge, taking the strength of the feature relationship between each of the feature parameter data as an edge weight, and constructing a directed weighted graph between each of the feature parameter data;
[0027] For each of the feature parameter data, according to the attention weight between the feature parameter data and the feature parameter data adjacent to it, the edge weight between the feature parameter data and the feature parameter data adjacent to it is updated; the updated edge weights are processed in descending order; according to the edge weights processed in descending order, the edges corresponding to the edge weights sorted after the preset position in the edge weights processed in descending order are deleted;
[0028] Determining the strength of the characteristic relationship between each of the characteristic parameter data according to the deleted edge weights corresponding to each of the characteristic parameter data;
[0029] The edge weight between the feature parameter data and its adjacent feature parameter data in the directed weighted graph is updated using the following expression:
[0030]
[0031] In the formula, represents the updated edge weight value of the l′th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; T i,u represents the set of messages transmitted from characteristic parameter data u in power equipment i to characteristic parameter data v of its adjacent connection at all times; represents the edge weight value of the l′-1th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; Represents the attention weight of the l′th edge between the feature parameter data u in the power equipment i and its adjacent feature parameter data v at time t.
[0032] Optionally, the health status levels include: critical defect, serious defect, general defect, sub-health, and health.
[0033] Optionally, based on the characteristic parameter data of the electric power equipment, using the Raindrop learning algorithm to solve the pre-built health assessment model of the electric power equipment to obtain the health status level of the electric power equipment includes:
[0034] Performing high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data;
[0035] Classifying the representative vectors of the characteristic parameter data according to time to obtain representative vectors of the characteristic parameter data at the same time and representative vectors of the characteristic parameter data at different times;
[0036] A graph neural network is used to aggregate the representation vectors of the feature parameter data at the same time to obtain an aggregated feature vector at the same time;
[0037] Aggregating the representation vectors of the feature parameter data at different times using an attention mechanism to obtain aggregated feature vectors at different times;
[0038] Performing vector concatenation of the aggregated feature vectors at the same time and the aggregated feature vectors at different times to obtain a graph representation vector;
[0039] Inputting the graph representation vector into the cross entropy loss function of the health assessment model of the power equipment to obtain probability values corresponding to different health status levels of the power equipment;
[0040] The health status level of the power equipment is obtained according to the probability values corresponding to different health status levels of the power equipment.
[0041] Optionally, after performing high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data, and before classifying the representation vector of the characteristic parameter data according to time to obtain a representation vector of the characteristic parameter data at the same time and a representation vector of the characteristic parameter data at different times, the method further includes:
[0042] Determine whether there is missing characteristic parameter data in the power equipment, and when there is missing characteristic parameter data, obtain a representation vector of adjacent characteristic parameter data of the missing characteristic parameter data;
[0043] Calculating the attention weight between the missing feature parameter data and the adjacent feature parameter data;
[0044] According to the attention weight, obtaining a representation vector of the missing feature parameter data;
[0045] The calculation formula of the attention weight between the missing feature parameter data and the adjacent feature parameter data is as follows:
[0046]
[0047] In the formula, represents the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t; μ(·) represents the message transfer function; represents the representation vector of the characteristic parameter data u in the power equipment i at time t under the high-dimensional feature mapping; D represents the trainable weight matrix; r v A trainable weight vector representing the feature parameter data v; Represents the time feature representation vector of the feature parameter data u at time t; T represents transpose.
[0048] Based on the same inventive concept, the present invention also provides a data-knowledge fusion driven power equipment status assessment system, comprising:
[0049] Data acquisition module: used to obtain characteristic parameter data of power equipment;
[0050] A status assessment module is used to solve the pre-built health assessment model of the power equipment based on the characteristic parameter data of the power equipment using the Raindrop learning algorithm to obtain the health status level of the power equipment;
[0051] The health assessment model in the state assessment module is constructed based on the strongly labeled samples of the power equipment acquired in advance and the weakly labeled samples obtained by formally expressing the knowledge and experience of the power equipment.
[0052] Optionally, the health assessment model of the power equipment in the state assessment module includes the following construction process:
[0053] Obtain strong annotation samples and knowledge and experience of the power equipment;
[0054] Formalize the knowledge and experience to obtain weakly labeled samples;
[0055] According to the strongly labeled samples, construct a cross entropy loss function of the strongly labeled samples;
[0056] According to the weakly labeled samples, construct a knowledge regularization term of the weakly labeled samples;
[0057] Constructing a knowledge function according to the strongly labeled samples and the weakly labeled samples;
[0058] Obtaining a cross entropy loss function of the health assessment model of the power equipment according to the cross entropy loss function of the strongly labeled samples, the knowledge regularization term of the weakly labeled samples and the knowledge function;
[0059] The health assessment model of the power equipment is obtained with the goal of minimizing the cross entropy loss function of the health assessment model of the power equipment.
[0060] Optionally, the state evaluation module constructs a knowledge function according to the strongly labeled samples and the weakly labeled samples, including:
[0061] According to the strongly labeled samples and the weakly labeled samples, the strength of the characteristic relationship between the characteristic parameter data in the electric power equipment is obtained;
[0062] According to the strength of the characteristic relationship between the characteristic parameter data in the power equipment, a knowledge function corresponding to each set of characteristic relationships is constructed.
[0063] Optionally, the expression corresponding to the cross entropy loss function of the health assessment model in the state assessment module is as follows:
[0064]
[0065] Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ 1 represents a hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 Indicates the weight coefficient value corresponding to the first group of feature relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationship between characteristic parameter data α and characteristic parameter data β; l=1…n; n represents the number of groups of characteristic parameter data forming the characteristic relationship.
[0066] Optionally, the knowledge function E corresponding to the first set of characteristic relations between the characteristic parameter data α and the characteristic parameter data β in the state evaluation module is l The expression of (α,β) is as follows:
[0067]
[0068] Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,n}; n represents the number of groups of feature parameter data forming the feature relations; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data at time X; Ω G represents the sampling set of characteristic parameter data at time G; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value.
[0069] Optionally, after obtaining the strength of the characteristic relationship between each characteristic parameter data in the power equipment according to the strongly labeled samples and the weakly labeled samples, the state evaluation module further includes:
[0070] Taking each of the feature parameter data as a vertex, taking the feature relationship between each of the feature parameter data as an edge, taking the strength of the feature relationship between each of the feature parameter data as an edge weight, and constructing a directed weighted graph between each of the feature parameter data;
[0071] For each of the feature parameter data, according to the attention weight between the feature parameter data and the feature parameter data adjacent to it, the edge weight between the feature parameter data and the feature parameter data adjacent to it is updated; the updated edge weights are processed in descending order; according to the edge weights processed in descending order, the edges corresponding to the edge weights sorted after the preset position in the edge weights processed in descending order are deleted;
[0072] Determining the strength of the characteristic relationship between each of the characteristic parameter data according to the deleted edge weights corresponding to each of the characteristic parameter data;
[0073] The edge weight between the feature parameter data and its adjacent feature parameter data in the directed weighted graph is updated using the following expression:
[0074]
[0075] In the formula, represents the updated edge weight value of the l′th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; T i,u represents the set of messages transmitted from characteristic parameter data u in power equipment i to characteristic parameter data v of its adjacent connection at all times; represents the edge weight value of the l′-1th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; Represents the attention weight of the l′th edge between the feature parameter data u in the power equipment i and its adjacent feature parameter data v at time t.
[0076] Optionally, the health status levels in the status assessment module include: critical defects, serious defects, general defects, sub-health, and health.
[0077] Optionally, the status assessment module is specifically used to:
[0078] Performing high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data;
[0079] Classifying the representative vectors of the characteristic parameter data according to time to obtain representative vectors of the characteristic parameter data at the same time and representative vectors of the characteristic parameter data at different times;
[0080] A graph neural network is used to aggregate the representation vectors of the feature parameter data at the same time to obtain an aggregated feature vector at the same time;
[0081] Aggregating the representation vectors of the feature parameter data at different times using an attention mechanism to obtain aggregated feature vectors at different times;
[0082] Performing vector concatenation of the aggregated feature vectors at the same time and the aggregated feature vectors at different times to obtain a graph representation vector;
[0083] Inputting the graph representation vector into the cross entropy loss function of the health assessment model of the power equipment to obtain probability values corresponding to different health status levels of the power equipment;
[0084] The health status level of the power equipment is obtained according to the probability values corresponding to different health status levels of the power equipment.
[0085] Optionally, after the state assessment module performs high-dimensional feature mapping on the characteristic parameter data of the electric equipment to obtain a representation vector of the characteristic parameter data, and before classifying the representation vector of the characteristic parameter data according to time to obtain a representation vector of the characteristic parameter data at the same time and a representation vector of the characteristic parameter data at different times, the method further includes:
[0086] Determine whether there is missing characteristic parameter data in the power equipment, and when there is missing characteristic parameter data, obtain a representation vector of adjacent characteristic parameter data of the missing characteristic parameter data;
[0087] Calculating the attention weight between the missing feature parameter data and the adjacent feature parameter data;
[0088] According to the attention weight, obtaining a representation vector of the missing feature parameter data;
[0089] The calculation formula of the attention weight between the missing feature parameter data and the adjacent feature parameter data is as follows:
[0090]
[0091] In the formula, represents the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t; μ(·) represents the message transfer function; represents the representation vector of the characteristic parameter data u in the power equipment i at time t under the high-dimensional feature mapping; D represents the trainable weight matrix; r v A trainable weight vector representing the feature parameter data v; represents the time characteristic representation vector of power equipment i at time t; T represents transpose.
[0092] Compared with the prior art, the present invention has the following beneficial effects:
[0093] The present invention provides a data-knowledge fusion driven power equipment status assessment method and system, comprising: acquiring characteristic parameter data of the power equipment; based on the characteristic parameter data of the power equipment, solving a pre-constructed health assessment model of the power equipment using a Raindrop learning algorithm to obtain a health status level of the power equipment; wherein the health assessment model is constructed based on weakly labeled samples obtained by formally expressing the strong labeled samples of the power equipment acquired in advance and the knowledge and experience of the power equipment; the present invention solves the health assessment model introducing weakly labeled samples through a Raindrop learning algorithm, which is beneficial to enhancing the effective information of the sample space, improving the generalization ability of the model, and can overcome the problem of reduced accuracy of power equipment health assessment caused by irregular characteristic parameter data, and can better adapt to the health assessment needs of actual power equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0094] Figure 1 A schematic diagram of a flow chart of a data-knowledge fusion driven power equipment status assessment method provided by the present invention;
[0095] Figure 2 A schematic diagram of the process execution of a health assessment model in a power equipment status assessment method driven by data-knowledge fusion provided by the present invention;
[0096] Figure 3 A schematic diagram of irregular data collection in a data-knowledge fusion driven power equipment status assessment method provided by the present invention;
[0097] Figure 4 A schematic diagram of a process of a power equipment status assessment method driven by data-knowledge fusion provided by the present invention using a health assessment model guided by intermediate knowledge based on the Raindrop learning algorithm;
[0098] Figure 5 A schematic diagram of a framework for feature classification of irregular data of power equipment in a data-knowledge fusion driven power equipment status assessment method provided by a specific embodiment of the present invention;
[0099] Figure 6 A schematic diagram of the structural composition of a data-knowledge fusion driven power equipment status assessment system provided for a specific embodiment of the present invention. DETAILED DESCRIPTION
[0100] The present invention proposes a data-knowledge fusion driven power equipment status assessment method and system. The specific implementation manner of the present invention is further described in detail below with reference to the accompanying drawings.
[0101] Embodiment 1:
[0102] A method for learning the characteristics of power equipment status data, the flow chart is as follows Figure 1 As shown, including:
[0103] Step 1: Obtain characteristic parameter data of power equipment;
[0104] Step 2: Based on the characteristic parameter data of the power equipment, the Raindrop learning algorithm is used to solve the pre-built health assessment model of the power equipment to obtain the health status level of the power equipment;
[0105] Among them, the health assessment model is constructed based on the strongly labeled samples of power equipment acquired in advance and the weakly labeled samples obtained by formally expressing the knowledge and experience of power equipment.
[0106] Power equipment plays a vital role in the power system. They are designed for various links such as power generation, transmission, transformation and distribution. According to their functions and roles in the power system, power equipment can be mainly divided into two categories: power generation equipment and power supply equipment. Among them, power generation equipment can include: power station boilers, steam turbines, gas turbines, turbines, transformers, etc.; power supply equipment can include: transmission lines, mutual inductors, contactors, circuit breakers, disconnectors, fuses, power cables and insulators, etc. In general, based on ledger data, online monitoring data, operation and maintenance data, etc., the actual characteristic parameter data set of power equipment in terms of insulation, heat generation, discharge, etc. can be obtained. The characteristic parameter data of the power equipment in step 1 above refers to the key parameters that can reflect the status and performance of the power equipment. These parameters are crucial for evaluating the health status of the power equipment, monitoring the operating status, and performing fault diagnosis and preventive maintenance. The following are some common characteristic parameter data of power equipment: electrical parameter data, thermal parameter data, mechanical parameter data, chemical parameter data, insulation parameter data, control parameter data, equipment status data and environmental parameter data, etc. However, due to different sampling times of different characteristic parameter data of power equipment, the sampling interval of the same characteristic parameter data is not fixed. Therefore, the actual characteristic parameter data is irregularly distributed.
[0107] In order to truly reflect the health status of power equipment, it is necessary to establish a comprehensive, multi-level, systematic and scientific health status assessment index system, including but not limited to comprehensive performance in terms of insulation, heating, grounding, etc. The characteristic parameter data of power equipment can be divided into static characteristics and dynamic characteristics according to the characteristics. Static characteristics refer to the inherent properties of power equipment, which directly determine the characteristics and performance of power equipment in the working environment. Dynamic characteristics refer to the state characteristics of power equipment that change over time during operation, which can be divided into four aspects: electrical performance, physical and chemical performance, main body appearance and accessory performance, and surrounding environment. The present invention integrates the State Grid Corporation enterprise standards such as Q / GDW 645 "Guidelines for Distribution Network Equipment Status Evaluation", Q / GDW 644 "Guidelines for Distribution Network Equipment Status Maintenance" and "State Grid Transmission and Transformation Equipment Status Maintenance Test Procedure", as well as DL / T 596 "Preventive Test Procedure for Power Equipment" and other power industry related standards, and combines the research results of experts and scholars to construct a power equipment health status assessment index system, which specifically includes static characteristics and dynamic characteristics, and proposes specific indicators for each aspect.
[0108] In view of the irregularly distributed characteristic parameter data of power equipment, the present invention selects the Raindrop learning algorithm to evaluate the health status of power equipment, processes the characteristic parameters at the same time through the graph neural network, and processes the characteristic parameters at different times through the attention mechanism, so as to realize multi-classification learning directly using irregular data and complete the status evaluation of power equipment;
[0109] In one implementation, in the above step 2, based on the characteristic parameter data of the power equipment, the process of solving the pre-built health assessment model of the power equipment using the Raindrop learning algorithm to obtain the health status level of the power equipment may include:
[0110] Perform high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data;
[0111] Classifying the representation vectors of the characteristic parameter data according to time to obtain the representation vectors of the characteristic parameter data at the same time and the representation vectors of the characteristic parameter data at different times;
[0112] The graph neural network is used to aggregate the representation vectors of the feature parameter data at the same time to obtain the aggregated feature vector at the same time;
[0113] The attention mechanism is used to aggregate the representation vectors of feature parameter data at different times to obtain aggregated feature vectors at different times;
[0114] Concatenate the aggregated feature vectors at the same time and the aggregated feature vectors at different times to obtain a graph representation vector;
[0115] The graph representation vector is input into the cross entropy loss function of the health assessment model of the power equipment to obtain the probability values corresponding to different health status levels of the power equipment;
[0116] The health status level of the power equipment is obtained according to the probability values corresponding to different health status levels of the power equipment.
[0117] For example, the above health status levels may include: critical defect, serious defect, general defect, sub-health, and health; different health status levels are distinguished by the performance and expression of the power equipment.
[0118] For example, the output of the above health assessment model is f δ (x i ) indicates that when the model parameter is δ, according to the characteristic parameter data x of the power equipment i i Calculate the probability value Pr(y i =C|x i ; δ), where the health category with the highest probability is taken as the feature parameter data x i The final health status level. Characteristic parameter data x for power equipment i i , f δ (x i ) will output a 5-dimensional vector, representing the probability values of the corresponding 5 health status levels, where δ is the parameter of the model. The expression corresponding to the output of the health assessment model is as follows:
[0119]
[0120] Among them, f δ (x i ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δ i The corresponding output; when C = 1, Pr (y i =1|x i ; δ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δ i Corresponding health status level y i is the probability value corresponding to the critical defect; Pr(y i =2|x i ; δ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δ i Corresponding health status level y i is the probability value corresponding to serious defects; Pr(y i =3|x i ; δ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δi Corresponding health status level y i is the probability value corresponding to the general defect; Pr(y i =4|x i ; δ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δ i Corresponding health status level y i is the probability value corresponding to sub-health; Pr(y i =5|x i ; δ) represents the irregular characteristic parameter data x of the i-th power equipment when the model parameter is δ i Corresponding health status level y i is the probability value corresponding to health;
[0121] In this implementation, the Raindrop learning algorithm is actually a multi-classification learning method that directly utilizes irregular data through graph neural networks and attention mechanisms. In this problem, based on the characteristic relationship between the characteristic parameter data of power equipment, the Raindrop learning algorithm is used to construct a directed weighted graph. For example, the directed weighted graph of power equipment i can be represented as G i = {ν,ε i}, where ν is a node in a directed weighted graph, and is also used to represent feature parameter data ν, ε i represents an edge, describing the characteristic relationship between the characteristic parameter data related to the power equipment i. In a directed weighted graph, a directed edge from node u to node v is represented by a triple Edge{u,e i,uv ,v},e i,uv ∈[0,1], where e i,uv is the directed edge weight, which is used to indicate the strength of the characteristic relationship between the characteristic parameter data u of the power equipment i and its adjacent characteristic parameter data v. i,uv If it is equal to 0, it means that there is no message transmission between feature parameter data u and feature parameter data v, and the two feature parameter data are unrelated. It should be noted that the importance of feature parameter data u to feature parameter data v is different from the importance of feature parameter data v to feature parameter data u. i,uv ,v} represents the message passing mechanism between u and v. When node u has a sampled value, it will use e i,uv Sending a neural message to node v, that is, the relationship graph between feature parameter data can be regarded as a directed weighted graph.
[0122] In this implementation, the Raindrop learning algorithm first maps the non-empty feature parameter data u of the sampled power equipment i to a vector representation in a high-dimensional space, and then uses a directed weighted graph G i The directed edge triple Edge{u,ei,uv ,v} passes the message from the feature parameter data u to the feature parameter data v that is adjacent to it. Therefore, the feature parameter data u will not only affect its own representation vector, but also the representation vector of the adjacent feature parameter data v. Secondly, the Raindrop learning algorithm uses the time attention mechanism to aggregate the representation vector of each feature parameter data to obtain the representation vector of the feature parameter data at each time. By considering the feature parameter representation at different times, the constructed health assessment model can capture the trend and characteristics of the feature parameter data over time. Finally, the representation vector and time feature of the feature parameter data are spliced, and then the representation vector of all feature parameters is calculated in combination with the attention weight, which can be used for the health status assessment of power equipment.
[0123] In this implementation, after the characteristic parameter data of the electric power equipment is subjected to high-dimensional feature mapping to obtain the representation vector of the characteristic parameter data, the representation vector of the characteristic parameter data is classified according to time to obtain the representation vector of the characteristic parameter data at the same time and the representation vector of the characteristic parameter data at different times, the following may also be included:
[0124] Determine whether there is missing characteristic parameter data in the power equipment, and when there is missing characteristic parameter data, obtain a representation vector of adjacent characteristic parameter data of the missing characteristic parameter data;
[0125] Calculate the attention weights between the missing feature parameter data and the adjacent feature parameter data;
[0126] According to the attention weight, the representation vector of the missing feature parameter data is obtained;
[0127] Among them, the calculation formula of the attention weight between the missing feature parameter data and the adjacent feature parameter data is as follows:
[0128]
[0129] In the formula, represents the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t; μ(·) represents the message transfer function; represents the representation vector of the characteristic parameter data u in the power equipment i at time t under the high-dimensional feature mapping; r v Represents the trainable weight vector of feature parameter data v, r v ∈R dr , R dr represents a real vector of dr dimensions; Represents the time feature representation vector of the feature parameter data u at time t, R dtrepresents a real vector of dt dimension; D represents a trainable weight matrix, which is used to map the dh dimension vector to a (dt+dr) dimension vector; T represents transpose;
[0130] Since the feature parameter data at different times are i The importance of the vectors is different. The following formula can be used to calculate the attention weight β of the characteristic parameter data v of the power equipment i at each time: i,v :
[0131]
[0132] Among them, softmax(·) represents the softmax activation function; Q i,v The first intermediate matrix corresponding to the representation vector of the characteristic parameter data v of the electric power equipment i at each time; K′ i,v A second intermediate matrix corresponding to a representation vector representing characteristic parameter data v of electric power equipment i at each time; represents the trainable parameters of the first intermediate matrix; represents the trainable parameters of the second intermediate matrix; H i,v Represents the concatenated vector related to the power equipment i and the characteristic parameter data v; d k represents the dimension size of the second intermediate matrix; s represents the scaling factor of softmax normalization; then the characteristic parameter data v of power equipment i at time t is used to represent the vector under high-dimensional feature mapping With the time feature representation vector Concatenate to form a concatenated vector: Represents the time feature representation vector of power equipment i at time t; then the following calculation formula is used to combine the attention weight of the characteristic parameter data v of power equipment i at time t Calculate the representation vector Z of the characteristic parameter data v i,v :
[0133]
[0134] Among them, W 0 Represents the weight vector parameter matrix shared by all power devices; T i,v represents the set of messages transmitted from the characteristic parameter data v in the power device i to the characteristic parameter data u of its adjacent connection at all times; represents the attention weight of the characteristic parameter data v of the power equipment i at time t; The representation vector of the characteristic parameter data v in the power equipment i at time t under the high-dimensional feature mapping; represents the time characteristic representation vector of power equipment i at time t; finally, the representation vectors of all characteristic parameter data are concatenated using the following calculation formula to form the representation vector Z of power equipment i i For health assessment:
[0135]
[0136] Among them, Z i,1 A representation vector representing the first characteristic parameter data in the power device i; Indicates the Mth power device i 0 The representation vector of feature parameter data; M 0 It represents the total number of characteristic parameter data in power equipment i.
[0137] In this implementation, it is assumed that there is a directed weighted graph G i In the above, the characteristic parameter data u is represented by node u, and the characteristic parameter data v is represented by node v. Node u and node v are two adjacent connected nodes, but the sampling value of the characteristic parameter data corresponding to node v at time t is missing. At this time, the representation vector of the missing node v is obtained through the message passing of the Raindrop learning algorithm. The process specifically includes: first, the input of node u (that is, characteristic parameter data u) in the power equipment i at time t is mapped to the representation vector in the high-dimensional space as in dh represents the dimension of the vector, R dh represents a real matrix of dh dimension. Then along the directed weighted graph G i The directed edge Edge{u,e i,uv ,v} transmits the neural message from node u to the adjacent node v. Specifically, first, according to the attention weight between node u and node v in power device i at time t Then calculate the representation vector of node v in power device i at time t by the following formula:
[0138]
[0139] Where μ(·) represents the message passing function; Specifically represents the representation vector of node v in power equipment i under high-dimensional feature mapping at time t; u Represents a directed weighted graph G i The shared weight parameter of the middle node u; w v Represents a directed weighted graph G i The shared weight parameter of the middle node v; T represents transposition; e i.uv represents the weight of the directed edge between node u and node v in power equipment i;
[0140] In this implementation, the directed weighted graph Gi The edges connecting the feature parameter data in will also undergo updates and pruning, including:
[0141] Taking each feature parameter data as a vertex, taking the feature relationship between each feature parameter data as an edge, taking the strength of the feature relationship between each feature parameter data as an edge weight, a directed weighted graph between each feature parameter data is constructed;
[0142] For each feature parameter data, according to the attention weight between the feature parameter data and its adjacent feature parameter data, the edge weight between the feature parameter data and its adjacent feature parameter data is updated; the updated edge weights are processed in descending order; according to the edge weights processed in descending order, the edges corresponding to the edge weights sorted after the preset position in the edge weights processed in descending order are deleted;
[0143] Determine the strength of the feature relationship between each feature parameter data according to the deleted edge weight corresponding to each feature parameter data;
[0144] For example, the edge weight between the feature parameter data and its adjacent connected feature parameter data in the above directed weighted graph is updated using the following expression:
[0145]
[0146] In the formula, represents the updated edge weight value of the l′th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; T i,u represents the set of messages transmitted from characteristic parameter data u in power equipment i to characteristic parameter data v of its adjacent connection at all times; represents the edge weight value of the l′-1th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; Represents the attention weight of the l′th edge between the feature parameter data u in the power equipment i and its adjacent feature parameter data v at time t.
[0147] In the network architecture of the Raindrop learning algorithm, l′ layers of Raindrop networks are stacked to build the model. In the graph initialization stage, all feature parameter data are connected to each other, but this will cause incorrect connections to be established between some unrelated feature parameter data, making the correlation of feature parameter data inconsistent with the actual situation. Therefore, it is necessary to update the weights of the edges and prune the structure of the graph. Specifically, based on the above calculation formula: Obtain the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t, using the formula: Update the edge weight value of the l′th edge between the characteristic parameter data u in the power equipment i and its adjacent characteristic parameter data v, and prune unimportant edges. At each layer, sort the edge weights connected to each node, and only retain the preset K% edge weights; by updating the edge weights, the erroneous connections between the characteristic parameter data can be corrected so that the correlation conforms to the knowledge rules; by pruning the structure of the graph, unnecessary edges can be removed, thereby reducing the model complexity and improving the computational efficiency.
[0148] In addition, in the field of power equipment health assessment, the cost of obtaining labeled samples is high, resulting in fewer accurately labeled samples. At the same time, there are more power equipment in healthy and sub-healthy states, while there are relatively fewer power equipment in defective states, and there is an imbalance problem in the data. Therefore, it is difficult to design a health assessment model with sufficient generalization ability and robustness. However, a large amount of prior knowledge and rules have been accumulated in the field of power equipment health assessment. If the relevant knowledge and experience can be formally expressed and introduced into the learning process, it will be able to effectively cope with the challenges of insufficient and unbalanced samples, which is conducive to solving the above problems and improving the accuracy of the assessment results.
[0149] In one implementation, the health assessment model of the power equipment may include the following construction process:
[0150] Obtain strong annotation samples and knowledge and experience of power equipment;
[0151] Formalize knowledge and experience to obtain weakly labeled samples;
[0152] According to the strongly labeled samples, a cross entropy loss function of the strongly labeled samples is constructed;
[0153] According to the weakly labeled samples, a knowledge regularization term of the weakly labeled samples is constructed;
[0154] Construct a knowledge function based on strongly labeled samples and weakly labeled samples;
[0155] According to the cross entropy loss function of the strongly labeled samples, the knowledge regularization term and knowledge function of the weakly labeled samples, the cross entropy loss function of the health assessment model of the power equipment is obtained;
[0156] The goal is to minimize the cross entropy loss function of the health assessment model of power equipment, and obtain the health assessment model of power equipment.
[0157] In this implementation, the process of constructing the knowledge function based on the strongly labeled samples and the weakly labeled samples may include:
[0158] According to the strongly labeled samples and the weakly labeled samples, the strength of the characteristic relationship between the characteristic parameter data in the power equipment is obtained;
[0159] According to the strength of the characteristic relationship between the characteristic parameter data in the power equipment, the knowledge function corresponding to each set of characteristic relationships is constructed.
[0160] For example, the expression corresponding to the cross entropy loss function of the above health assessment model is as follows:
[0161]
[0162] Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ 1 represents a hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 Indicates the weight coefficient value corresponding to the first group of feature relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationships between the characteristic parameter data α and the characteristic parameter data β.
[0163] For example, the knowledge function E corresponding to the first set of characteristic relations between the characteristic parameter data α and the characteristic parameter data β is l The expression of (α,β) is as follows:
[0164]
[0165] Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,n}; n represents the number of groups of feature parameter data forming the feature relations; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data X at each time; Ω G represents the sampling set of characteristic parameter data G at each time; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value.
[0166] The schematic diagram of the knowledge-guided health assessment model based on the Raindrop learning algorithm proposed in this invention is as follows: Figure 2As shown in the figure. The collected irregular feature parameter data include winding insulation resistance, core grounding current, oil breakdown voltage, oil micro-water content, lead joint temperature, body appearance and load conditions and other multi-dimensional feature parameters. The knowledge part includes the introduction of weakly labeled samples and feature relationship guided learning. In the figure, [t 1 ,…,t T0 ] represents the time series. In terms of introducing weakly labeled samples, the original accurately labeled samples are used as strong labeled samples, that is, the time series in the figure is [t 2 ,…,t T0 ] represents a strongly labeled sample. For the strongly labeled sample, the feature parameter data u is converted into 3 Representation vector of high-dimensional feature map under time The edge weight e between feature parameter data u and v i.uv , the attention weight between feature parameter data u and v And the trainable weight vector r of the feature parameter data v v With t 3 The time feature representation vector of the feature parameter data u under time Concatenated vectors Get 3 The representation vector of the characteristic parameter data u of the power equipment i under high-dimensional feature mapping under time Forming the first intermediate matrix Q about the characteristic parameter data v i,v ; For unlabeled samples, the time series in the figure is t 1 The feature parameter data of represents weakly labeled samples, using existing knowledge and experience, such as using feature parameter data u in t 1 Representation vector of high-dimensional feature map under time To provide weak labeling information, the feature parameter data u is 1 Representation vector of high-dimensional feature map under time The edge weight e between feature parameter data u and v i.uv , the attention weight between feature parameter data u and v And the trainable weight vector r of the feature parameter data v v With t 1 The time feature representation vector of the feature parameter data u under time The stitching vector Get 1 The representation vector of the characteristic parameter data u of the power equipment i under high-dimensional feature mapping under time Forming a second intermediate matrix K′ about the characteristic parameter data v i,v, the first intermediate matrix and the second intermediate matrix are activated by Softmax function to obtain the attention weight β of the characteristic parameter data v of the power equipment i at each time i,v , and then according to the attention weight β i,v Calculate the representation vector Z of the characteristic parameter data v i,v , and finally the following graph is formed to represent the vector Z i For health assessment;
[0167]
[0168] Specifically, assuming that the original labeling result of power equipment i is y (i) =[0,1,0,0,0], i∈{1,2…M}, which means that in the strongly labeled samples containing M power equipment, power equipment i is in a serious defect state; while for power equipment j in the unlabeled samples, although the knowledge rules cannot directly mark its precise health state, they can indirectly mark its possible health level range. For example j∈{1,2…N}, which means that among the unlabeled samples containing N power equipment, the probability that power equipment j is a weakly labeled sample and has a critical defect or a serious defect is low.
[0169] Traditional machine learning methods can directly use strongly labeled samples, however, weakly labeled samples obtained by knowledge-guided learning strategies cannot be directly used by machine learning methods. Therefore, the present invention constructs a knowledge regularization term in the cross entropy loss function, which adds weakly labeled information to the loss function and transfers knowledge and experience to the model through back propagation to guide the model's learning and reasoning of unknown situations, thereby improving the performance and effect of the model.
[0170] The existing knowledge regularization term construction method of the loss function is expressed as follows:
[0171]
[0172] Among them, L CE (θ) represents the cross entropy loss function with respect to parameter θ; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ 1represents a custom hyperparameter. When its value is small enough, L CE (θ) is equivalent to the original loss function, that is, only strong labeled samples are used for training. 1 When the value is large, the knowledge regularization term will cause problems such as gradient vanishing or explosion during the training process, making it difficult for the model to learn and converge effectively. Therefore, when defining λ 1 It is necessary to carefully balance accuracy and stability and make appropriate adjustments to achieve excellent model performance.
[0173] When conducting feature relationship guided learning, the role is to strengthen the model's ability to understand and analyze the feature relationship between feature parameter data, so as to guide the learner to train towards optimization. The present invention utilizes the expert knowledge, industry standards and historical experience in the existing field of power equipment health assessment to mine and summarize the relationship between feature parameter data, establish a relationship model between feature parameter data, and divide the feature relationship between feature parameter data into strong correlation and weak correlation.
[0174] For the strong correlation between feature parameter data, for example, ε is used to represent a custom strong correlation parameter, and the value set is N 0 Indicates the number of groups of strongly correlated feature relationships, such as ε 1 Indicates the strong correlation parameter corresponding to the first group of feature relationships, Indicates the Nth 0 Strongly correlated parameters corresponding to group characteristic relationships. For example, the health status of transformer oil can be characterized by multidimensional characteristic parameter data. In general, the mechanism of action between characteristic parameter data is: the increase in the water content W, impurity content I and oil acidity A in the oil, as well as the increase in oil temperature U, will lead to a decrease in the insulation performance of the oil, which in turn causes the oil breakdown voltage V to decrease. The increase in impurity content I and water content W in the oil will cause the antioxidant content in the oil to decrease, leading to the oxidation reaction of the oil, thereby accelerating the acidification process of the oil. The increase in water content W, impurity content I and oil acidity A in the oil causes the thermal stability of the oil to decrease, and the oil temperature further increases. At this time, the increase in oil temperature will accelerate the decomposition and aging process of the oil. The characteristic relationship between characteristic parameter data is shown in the following formula:
[0175]
[0176] Among them, (e i,αβ &e i,βα )≥ε indicates that the strong correlation parameter value between the characteristic parameter data α and the characteristic parameter data β of the power equipment i is greater than or equal to ε, that is, there is a strong correlation between the characteristic parameter data α and the characteristic parameter data β; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; ei,βα represents the edge weight between the characteristic parameter data β and the characteristic parameter data α of the i-th power equipment; Ω represents the sampling set of the corresponding characteristic parameter data at any time; for example, Ω W represents the sampling set of the water content W in the oil at any time; Ω I represents the sampling set of impurity content I at any time; Ω A represents the sampling set of oil acidity A at any time; Ω U Represents the sampling set of oil temperature U at any time; Ω V represents the sampling set of oil breakdown voltage V at any time;
[0177] For the weak correlation between feature parameter data, for example, σ is used to represent a custom weak correlation parameter, and the value set is N 1 The number of groups representing weakly correlated feature relationships, such as σ 1 Indicates the weak correlation parameter corresponding to the first group of feature relationships, Indicates the Nth 1 Taking transformer as an example, in the gas chromatography research on insulating oil, the results show that when there are defects in transformer insulation, it will cause H 2 and abnormal production of hydrocarbon gas. When the temperature at the fault point gradually rises, the oil will undergo cracking and 4 , C 2 H 6 , C 2 H 4 , C 2 H 2 The three ratio method is widely used in the transformer fault diagnosis process. Its principle is to use the interdependence between the relative concentration of the gas components produced by the decomposition of the internal insulation material of the transformer under fault and the temperature. When selecting the characteristic gas, two gases with similar solubility and diffusion coefficient are selected from the five gases to form three comparison values for diagnosis. Therefore, the three comparison values C 2 H 2 With C 2 H 4 , CH 4 With H 2 , C 2 H 6 With C 2 H 4 The strong correlation between them is shown as follows:
[0178]
[0179] On this basis, other gases such as H 2 With C2 H 6 The contrast values of the same form a weak correlation, as shown in the following formula:
[0180]
[0181] Among them, (e i,αβ &e i,βα )≥ε indicates that the characteristic parameter data α and the characteristic parameter data β of the power equipment i are strongly correlated; (e i,αβ &e i,βα )<σ means that the weak correlation parameter value between the characteristic parameter data α and the characteristic parameter data β of the power equipment i is less than σ, that is, the characteristic parameter data α and the characteristic parameter data β are weakly correlated; Represents C 2 H 2 With C 2 H 4 The sampling collection at various times; Indicates CH 4 With H 2 The sampling collection at various times; Represents C 2 H 6 With C 2 H 4 A collection of samples at various times.
[0182] In summary, the knowledge-guided health assessment problem of power equipment based on the Raindrop learning algorithm can be modeled as an M1 problem. Specifically, the health assessment model is as follows: The minimization of the cross entropy loss function after weak annotation is introduced as the objective function, and it satisfies the formula:
[0183] and The knowledge constraints given in . Compared with the general method, the knowledge-guided health assessment method based on the Raindrop learning algorithm adopted in the present invention introduces knowledge regularization terms and knowledge condition constraints, and the loss function of the model is more complex and more difficult to converge. Therefore, based on the irregular data of power equipment, a knowledge-guided learning strategy based on the Raindrop learning algorithm is proposed to solve the above problem.
[0184] In the knowledge-guided learning strategy based on the Raindrop learning algorithm, in order to introduce weakly labeled samples, the knowledge rules are first expressed in a formal way. The present invention uses the representation form of propositional logic to make judgments based on knowledge rules and deduce possible conclusions. Assume A 0 Indicates the premise, B 0 Indicates the conclusion, S0 Representing the possible solution set, we can get the following formal representation:
[0185]
[0186] Among them, k represents a knowledge rule, and the arrow represents "defined as". When multiple knowledge rules coexist, they can be further represented by relevant logical or mathematical operation symbols.
[0187] Let K a ={k m (x)}, m=1…N a , K a Indicates that it contains N a A collection of knowledge rules, where x represents feature parameter data, k m (x) represents the mth knowledge rule, which is used to determine the possible health status level of power equipment i. Taking the transformer as an example, the knowledge rule of the weakly labeled sample can be expressed as:
[0188] 1) k 1 : {If winding insulation resistance>η 1 ||Core insulation resistance>η 2 then Pr(y>4)≤0}
[0189] 2) k 2 : {If H 2 >η 3 ||CH 4 >η 4 ||C 2 H 6 >η 5 ||C 2 H 4 >η 6 ||C 2 H 2 >η 7 then Pr(y>3)≤0}
[0190] 3) k 3 : {If the water content in oil>η 8 ||Flash point>η 9 ||Water soluble acid value <η 10 then Pr(y>2)≤0}
[0191] Among them, η 1 Indicates the associated threshold value of the winding insulation resistance; η 2 Indicates the threshold value of the core insulation resistance; η 3 Indicates H 2The correlation threshold of 4 Indicates CH 4 The correlation threshold of 5 Represents C 2 H 6 The correlation threshold of 6 Represents C 2 H 4 The correlation threshold of 7 Represents C 2 H 2 The correlation threshold of 8 Indicates the correlation threshold of the trace water content in oil; η 9 represents the associated threshold of the flash point; η 10 represents the associated threshold of water-soluble acid value; y represents the health status level;
[0192] Knowledge constraints pose a great challenge to feature relationship guided learning. If the knowledge constraints of the model M1 problem are analyzed and calculated during the iteration process, the efficiency of health assessment will be significantly reduced. The present invention changes the idea and converts it into the corresponding knowledge function according to the structure of the problem and the inequality characteristics of the constraints, and integrates it into the edge weight optimization process of the Raindrop learning algorithm. Specifically, the formed knowledge function set is represented as: Q = {E l (α,β)},l=1,2,3;E l (α, β) represents the knowledge function corresponding to the first set of characteristic relations between characteristic parameter data α and characteristic parameter data β. This function corresponds to a specific constraint term, such as E 1 The corresponding constraints are:
[0193]
[0194] E 2 The corresponding constraints are:
[0195]
[0196] E 3 The corresponding constraints are:
[0197]
[0198] That is, the constraint term expressed above can be expressed as e i,αβ ≥ε or e i,αβ <σ. Among them, e i,αβ is the left side of the constraint term, representing the edge weight between the feature parameter data, ε and σ are the right sides of the constraint term, representing the strong correlation parameter and weak correlation parameter of the feature parameter data respectively. The conditional constraint is converted into a knowledge function in the form of:
[0199]
[0200] Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,3}; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data X at each time; Ω G Represents the sampling set of feature parameter data G at each time; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value. In the process of feature relationship guided learning, the graph neural network in the Raindrop learning algorithm will automatically update and prune the edge weights between feature parameters. If the edge weight exceeds the corresponding parameter, the knowledge function will generate a penalty function to enhance the model's understanding and learning ability of the feature relationship of feature parameter data. Specifically, the knowledge function corresponding to the conditional constraint is expressed by the following expression:
[0201]
[0202] Among them, E 1 (α, β) represents the knowledge function corresponding to the first set of characteristic relations between characteristic parameter data α and characteristic parameter data β in power equipment i; in the first set of characteristic relations, characteristic parameter data α and characteristic parameter data β both belong to the sampling set Ω corresponding to transformer oil at any time I,V,A,U,W ;Ω G represents the sampling set of characteristic parameter data G at each time; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; ε 1 represents the strong correlation parameter corresponding to the first group of feature relations; M represents the number of power equipment in the strongly labeled samples; N represents the number of power equipment in the weakly labeled samples; E 2 (α, β) represents the knowledge function corresponding to the second set of characteristic relations between characteristic parameter data α and characteristic parameter data β; in the second set of characteristic relations, characteristic parameter data α and characteristic parameter data β both belong to the sampling set corresponding to the insulating oil at any time. ε 2 Indicates the strong correlation parameter corresponding to the second group of feature relationships; E 3 (α, β) represents the knowledge function corresponding to the third set of characteristic relations between characteristic parameter data α and characteristic parameter data β; in the second set of characteristic relations, characteristic parameter data α and characteristic parameter data β belong to the sampling set Ω, and do not belong to the sampling set corresponding to the insulating oil at any time. σ 1 Indicates the weak correlation parameter corresponding to the third group of characteristic relationships;
[0203] For different constraints, the present invention establishes corresponding knowledge functions respectively. Therefore, the expression corresponding to the cross entropy loss function of the health assessment model constructed by the present invention is as follows:
[0204]
[0205] Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ 1 represents a hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 Indicates the weight coefficient value corresponding to the first group of feature relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationships between the characteristic parameter data α and the characteristic parameter data β.
[0206] In order to solve the problem that the feature parameter set required for the existing power equipment health status assessment is mainly obtained by sampling through manual inspection or online monitoring, the sampling time interval of manual inspection is not fixed, and there are problems such as feature parameter measurement errors or transmission losses in online monitoring, and the actual data is irregular and there are few accurately labeled samples, which leads to the problem that the accuracy of the power equipment health assessment method is reduced, the present invention proposes a data-knowledge fusion driven power equipment status assessment method, by applying the Raindrop learning algorithm with high performance for processing incomplete and irregular data sets in the field of power equipment health assessment for the first time, weakly annotating the imprecisely labeled samples through knowledge, and introducing formally expressed domain knowledge rules in the model training process, using knowledge regularization terms to improve the loss function, and constructing a knowledge function to guide the edge weight parameter update process in the edge weight update process, thereby enhancing the effective information of the sample space and guiding the learner to optimize training. The above improvements can greatly improve the generalization ability of the model, and can overcome the challenges brought by irregular, missing and abnormal feature parameters, and can better adapt to the health assessment needs of actual power equipment: Raindrop learning algorithm integrates the advantages of graph neural network and attention mechanism, and effectively overcomes the challenges brought by a large number of missing values and abnormal values in irregular data by learning the relationship between nodes and edges in the topological structure and considering the characteristic relationship between feature parameter data at different times. For the task of health assessment of power equipment, Raindrop learning algorithm can show a higher health assessment accuracy rate than other baseline methods, and shows better performance for the identification of important defect states of power equipment. In addition, the present invention uses formalized knowledge rules to weakly annotate the original feature parameter data set, which can overcome the problems of high data acquisition cost and low utilization efficiency in health assessment tasks, enhance the effective information of sample space, and help improve the accuracy and robustness of the model. By constructing a learning strategy of knowledge function to guide the update process of Raindrop's edge weight parameters, the model's ability to understand and analyze the correlation between feature parameters can be enhanced, and accurate representation and guided learning of knowledge and experience can be achieved to improve the computational efficiency and performance effect of the algorithm, and effectively improve the performance and effect of the model.
[0207] Embodiment 2:
[0208] A specific embodiment is used to illustrate the specific process of a data-knowledge fusion-driven power equipment status assessment method provided by the present invention for the problems of non-fixed sampling intervals of characteristic parameters, incorrect collection of some parameters, or transmission loss in actual power equipment health assessment: Taking the 110kV transformer health status assessment as an example, firstly, a 110kV power equipment health assessment index system is constructed and the health status level is divided; then, based on the Raindrop method, non-precisely labeled samples are weakly labeled through knowledge, and the knowledge regularization term is used to improve the loss function; in addition, in the edge weight update process, a knowledge function is constructed to guide the edge weight parameter update process. By guiding the learning strategy of Raindrop through knowledge, the generalization ability of the model can be improved, and the challenges brought by irregular, missing, and abnormal characteristic parameters can be effectively overcome. The following key steps are included:
[0209] (1) Construct a transformer health assessment index system and classify the health status levels
[0210] For the irregular data of 110kV transformer, Figure 3 As shown, it includes ledger data, live detection data, online monitoring data, maintenance and defect record data, oil test data, electrical test data, bushing test data and short-circuit no-load test data, and takes commissioning date, insulation resistance, load condition, oil chromatogram, main body appearance, oil temperature and oil level as inspection items. The benchmark period is March, and the time series is used. Indicates; oil dissolved gas analysis and bushing test are routine test items, with a reference period of 1 year or 3 years respectively; and when diagnosing a transformer fault, short-circuit no-load test, induction withstand voltage or partial discharge measurement and other diagnostic tests should be carried out according to the situation to find the cause of the fault, such as Figure 4 As shown in the figure, the whole process includes the introduction of weakly labeled samples and feature relationship guided learning, which specifically includes: formalizing the knowledge and experience of power equipment to obtain weakly labeled samples; constructing the cross entropy loss function of the strongly labeled samples according to the strongly labeled samples; constructing the knowledge regularization term of the weakly labeled samples according to the weakly labeled samples; constructing the knowledge function according to the strongly labeled samples and the weakly labeled samples, for example, according to the water content W in the oil, the impurity content I, the oil acidity A, the oil temperature U, the oil breakdown voltage V and the C in the transformer oil 2 H 2 With C 2 H 4 , CH 4 With H 2 , C 2 H 6 With C 2 H 4 The strength of the characteristic relationship between them is used to jointly construct the knowledge function; finally, the health assessment model of power equipment is constructed based on the cross entropy loss function of the strongly labeled samples, the knowledge regularization term of the weakly labeled samples and the knowledge function;
[0211] The goal is to minimize the cross entropy loss function of the health assessment model of power equipment, and obtain the health assessment model of power equipment. For example, in terms of electrical performance, the characteristic quantities such as winding insulation resistance, core insulation resistance and winding DC resistance are considered for the electrical test and online monitoring of the transformer. In terms of physical and chemical performance, the characteristic quantities such as the micro-water content in the oil, the furfural content level and the water-soluble acid value are considered for the state of the transformer's insulating oil, insulating paper and insulating structure.
[0212] The transformer health assessment index system includes static indicators and dynamic indicators, such as Figure 5 As shown. Static characteristics refer to the inherent properties of the transformer, such as commissioning date, rated capacity, installation location and use environment; dynamic characteristics refer to the state characteristics of power equipment that change over time during operation, which can be divided into electrical performance, physical and chemical performance, main body appearance and accessory performance and surrounding environment; in dynamic characteristics, electrical performance can generally include: winding DC resistance, winding insulation resistance, bushing dielectric loss, bushing insulation resistance, load condition, load loss and three-phase imbalance; physical and chemical properties can include: oil temperature, flash point, volume resistivity, water-soluble acid value, micro-water content in oil, lead joint temperature; main body appearance, that is, accessory performance, generally includes: sealing condition, bushing appearance, main body appearance, lead joint temperature, etc.; surrounding environment generally includes: air temperature, rainfall and snowfall, etc.; the main factors affecting the health status of the transformer are the temperature that the insulation withstands, the water content in the oil-paper insulation, and the oxygen content in the oil. During the deterioration process of the transformer, as the operating temperature increases, the cellulose in the fiber paperboard will be thermally degraded to produce water, acidic substances and gases, which will further accelerate the deterioration process of the power equipment. Electrical tests and insulating oil tests can effectively obtain key characteristic quantities of electrical properties and physical and chemical properties, and are an important way to evaluate the health status of transformers. At the same time, the metal inside and outside the transformer will produce substances such as oxide scale or rust during the aging process. These substances will cause a small potential difference between the internal and external metal surfaces of the transformer, causing the metal to deteriorate faster, so it is also particularly important to monitor the status of each component of the transformer. In addition, most transformers are exposed to the outdoors for a long time, and are subjected to intermittent power load shocks and external forces such as dust, lightning strikes, strong winds, rain, snow and ice. Therefore, the internal and external environment, load conditions, etc. of the transformer should be monitored to comprehensively evaluate the health status of the transformer.
[0213] The health status classification of transformers is shown in Table 1. The transformer health assessment results and the corresponding health classification include five levels: healthy, sub-healthy, general defect, serious defect and critical defect.
[0214] Table 1 Classification of health levels of power equipment
[0215]
[0216] (2) Data-knowledge fusion driven power equipment status data feature learning method
[0217] The transformer state data feature learning method driven by data-knowledge fusion is based on the Raindrop learning algorithm, which introduces weakly labeled samples and feature relationships to guide learning:
[0218] For the introduction of weakly labeled samples, the knowledge rules of weakly labeled samples can be expressed as:
[0219] 1) k 1 : {If winding insulation resistance>η 1 ||Core insulation resistance>η 2 then Pr(y>4)≤0}
[0220] 2) k 2 : {If H 2 >η 3 ||CH 4 >η 4 ||C 2 H 6 >η 5 ||C 2 H 4 >η 6 ||C 2 H 2 >η 7 then Pr(y>3)≤0}
[0221] 3) k 3 : {If the water content in oil>η 8 ||Flash point>η 9 ||Water soluble acid value <η 10 then Pr(y>2)≤0}
[0222] Among them, η 1 —η 10 represents the threshold value, and Pr represents the probability. 1)k 1 The knowledge rule of means that when the characteristic parameter winding insulation resistance or core insulation resistance that characterizes the insulation structure strength of the transformer is higher than the threshold, the probability that it is in a healthy state is extremely low, and the weak labeling result is 2) k 2 The knowledge rule is that when the characteristic parameter of the gas generated by the decomposition of the transformer is higher than the threshold, the probability of it being in a healthy or sub-healthy state is extremely low. The weak labeling result is =[0.33,0.33,0.33,0,0]. 3)k 3The knowledge rule is that when the characteristic parameter characterizing the insulation strength of transformer oil is higher than the threshold, the probability that it is in a critical defect or serious defect state is extremely high. The weak labeling result is These knowledge rules indicate that different feature dimensions and indicators have different importance in the transformer health assessment process.
[0223] For feature relationship guided learning, the health status of transformer oil can be characterized by multidimensional feature parameter data. The mechanism of action between the feature parameter data is: the increase of the water content W, impurity content I and oil acidity A in the oil, as well as the increase of oil temperature U, will lead to a decrease in the insulation performance of the oil, and then cause the oil breakdown voltage V to decrease. The increase of impurity content I and water content W in the oil will cause the content of antioxidants in the oil to decrease, leading to the oxidation reaction of the oil, and then accelerate the acidification process of the oil. The increase of water content W, impurity content I and oil acidity A in the oil causes the thermal stability inside the oil to decrease, and the oil temperature to further increase. At this time, the increase in oil temperature will accelerate the decomposition and aging process of the oil. Due to the feature relationship guided learning, knowledge constraints bring great challenges. If the knowledge constraints of model M1 in the iterative process are analyzed and calculated, the efficiency of health assessment will be significantly reduced. The present invention converts the idea, according to the structure of the problem and the inequality characteristics of the constraints, it is converted into the corresponding knowledge function, and integrated into the edge weight optimization process of Raindrop. Specifically, the formed knowledge function set is expressed as Q={E l (α,β)},l=1,2,3,E l (α, β) represents the lth knowledge function.
[0224] (3) Train the transformer health status classifier and save the weight file
[0225] The transformer irregular data includes 235 groups of strongly labeled samples and 45 groups of weakly labeled samples, and the training set, validation set and test set are divided into 8:1:1 ratio. The characteristic parameter data of the transformer can be divided into static characteristic parameter data and dynamic characteristic parameter data according to the characteristics. The static characteristics can include 4 dimensions, including commissioning date, rated capacity, installation location and cooling method; the dynamic characteristics can include oil temperature and oil level, winding DC resistance, winding insulation resistance, core grounding current, capacitive bushing dielectric loss, oil breakdown voltage, oil micro-water content and tap changer performance, etc., a total of 40 dimensions.
[0226] For transformer irregular data, the present invention selects the existing GRU-D (Gated Recurrent Unit with Decay, gated recurrent unit with decay mechanism), BRITS (Bidirectional Recurrent Imputation for Time Series, bidirectional recurrent filling model for time series), mTAND (Multi-time Attention Networks) and Raindrop learning algorithm as baseline methods, and improves the Raindrop method with knowledge-guided learning, and obtains the Raindrop method (Weakly Labeled Raindrop, WL-Raindrop) that introduces weakly labeled samples and the Raindrop method (Weakly Labeled and Feature Relationship Guided Raindrop, WF-Raindrop) that introduces weakly labeled samples and feature relationship guided learning. These methods use different models to learn and reason about irregular data, as shown in Table 2. Among them, WL-Raindrop and WF-Raindrop belong to knowledge-guided Raindrop methods. Compared with the baseline method, they use knowledge-guided learning strategies in the training process to further improve performance. Among them, the GRU-D method uses gated recurrent units and attenuation mechanisms to learn irregular data features; the BRITS method uses bidirectional recurrent neural networks to learn missing information of feature parameters. The mTAND method uses an attention mechanism to capture the temporal correlation of irregular data and generate a temporal representation vector.
[0227] Table 2 Comparison of various health assessment methods
[0228]
[0229] The health assessment results of the baseline methods are shown in Table 3, where Accuracy is used to measure the proportion of samples correctly classified by the model to the total number of samples; Precision is used to measure the proportion of the correct number of instances predicted as positive by the model; Recall is used to measure the ability of the model to identify all actual positive classes; It can be seen from the table that the performance and performance of the Raindrop learning algorithm are significantly superior to other baseline methods. Specifically, in terms of accuracy, the Raindrop learning algorithm is 2.5% higher than the BRITS method, which performs better among other baseline methods; in terms of F1 score, it is 2.3% higher than the GRU-D method. The reason is that the Raindrop learning algorithm integrates the advantages of graph neural networks and attention mechanisms. By learning the relationship between nodes and edges in the topological structure and considering the correlation between feature parameters at different times, it effectively overcomes the challenges brought by a large number of missing values and outliers in irregular data. For the transformer health assessment task, the Raindrop learning algorithm shows a higher health assessment accuracy than other baseline methods, and shows good performance in identifying important defect states of transformers. However, the Accuracy and F1 score results of the Raindrop learning algorithm are only 83.2% and 79.5%. Therefore, the relevant knowledge and experience can be formalized and introduced into the learning process to further improve the performance and effect of the model.
[0230] Table 3 Health assessment results of baseline methods
[0231]
[0232] In this embodiment, the health assessment results of the knowledge-guided learning method based on the Raindrop learning algorithm are shown in Table 4. For the health assessment results of strongly labeled samples, the Accuracy and F1score indicators of WL-Raindrop and WF-Raindrop are higher than those of the basic Raindrop learning algorithm. The reason is that the original parameter set is weakly labeled using formally represented knowledge rules, which can overcome the problems of high data acquisition cost and low utilization efficiency in health assessment tasks, enhance the effective information of the sample space, and help improve the accuracy and robustness of the model. For the health assessment results of weakly labeled samples, the Accuracy and F1 score indicators of WF-Raindrop are 4.3% and 5.2% higher than those of WL-Raindrop. It can be seen that WF-Raindrop can enhance the model's understanding and analysis of the correlation between feature parameters by constructing a knowledge function to guide the update process of Raindrop's edge weight parameters, thereby effectively improving the performance and effect of the model.
[0233] Table 4 Health assessment results of the knowledge-guided Raindrop method
[0234]
[0235]
[0236] Since a corresponding weight file will be generated each time a training is completed during the training process, this embodiment selects to train the network 2000 times. When the health assessment result is better, the weight file of each training will be updated. When the loss value is small and tends to be stable, the network training is completed. And the performance evaluation is performed on the validation set, and the weight file is saved when the best model performance is achieved. To this end, in this embodiment, the weight file is saved when WF-Raindrop achieves the best model performance, and then the saved weight file is placed in the prediction network to achieve the evaluation of the health status of the transformer. Due to the problems of insufficient and unbalanced samples in the field of health assessment of power equipment, the present invention uses a data enhancement method to improve the robustness of the model in a complex environment. In the irregular data of power equipment, the number of samples is expanded by 5 times by deleting, adding some feature parameters, and introducing noise. This is not only conducive to increasing the data sample size and balancing data distribution, but also can reduce the risk of overfitting and improve the generalization ability of the model. The health status assessment of power equipment belongs to a multi-classification learning task in machine learning. In order to examine the accuracy of the health assessment results, this embodiment uses the classification accuracy (Accuracy) to evaluate the model's ability to classify power equipment in different health states. Since the health status of power equipment is similar to human health, most power equipment is in a healthy or sub-healthy state, and the importance of different health levels is also different. To this end, it is necessary to consider the precision (Precision) and recall (Recall) of the health assessment, as well as the weighted average (F1 score) of the precision (Accuracy) and recall rate. In addition, the role of health assessment is to accurately identify power equipment with defects, so as to take active operation and maintenance methods to achieve distribution asset management, so it is also necessary to give more weight to the defective status of power equipment.
[0237] Through continuous training, the loss value of the network will change. When the network loss value continues to decrease and tends to be stable, it can be considered that the network has converged. In order to ensure the effectiveness of the training, more than 200 trainings are generally performed. The information finally contained in the loss value is mainly: the cross entropy loss between the predicted result of the health status of the power equipment and the actual health status of the power equipment, thereby obtaining the network training weight. During the training process, a corresponding weight file will be generated for each completed training. The present invention chooses to train the network 2000 times. When the health assessment result is better, the weight file of each training will be updated. When the loss value is small and tends to be stable, the network training is completed. And the performance evaluation is performed on the validation set, and the weight file is saved when the best model performance is achieved. The saved weight file is placed in the prediction network to evaluate the health status of the power equipment.
[0238] This embodiment conducts health assessment in multiple scenarios based on the actual acquired 110kV transformer data, verifying that the proposed method can overcome the challenges brought by the loss of characteristic parameters. Similarly, for other power equipment such as circuit breakers, transmission lines, capacitors, etc., the Raindrop method can also be used in combination with knowledge and experience to directly use irregular data for power equipment health assessment.
[0239] Embodiment 3:
[0240] The present invention based on the same inventive concept also provides a data-knowledge fusion driven power equipment status assessment system, the structural composition diagram is as follows Figure 6 As shown, including:
[0241] Data acquisition module: used to obtain characteristic parameter data of power equipment;
[0242] Status assessment module: It is used to solve the pre-built health assessment model of power equipment based on the characteristic parameter data of power equipment using the Raindrop learning algorithm to obtain the health status level of the power equipment;
[0243] Among them, the health assessment model in the status assessment module is constructed based on the strong labeled samples of the power equipment acquired in advance and the weak labeled samples obtained by formally expressing the knowledge and experience of the power equipment.
[0244] The health assessment model of the power equipment in the above-mentioned status assessment module includes the following construction process:
[0245] Obtain strong annotation samples and knowledge and experience of power equipment;
[0246] Formalize knowledge and experience to obtain weakly labeled samples;
[0247] According to the strongly labeled samples, a cross entropy loss function of the strongly labeled samples is constructed;
[0248] According to the weakly labeled samples, a knowledge regularization term of the weakly labeled samples is constructed;
[0249] Construct a knowledge function based on strongly labeled samples and weakly labeled samples;
[0250] According to the cross entropy loss function of the strongly labeled samples, the knowledge regularization term and knowledge function of the weakly labeled samples, the cross entropy loss function of the health assessment model of the power equipment is obtained;
[0251] The goal is to minimize the cross entropy loss function of the health assessment model of power equipment, and obtain the health assessment model of power equipment.
[0252] The above-mentioned state evaluation module constructs a knowledge function based on the strongly labeled samples and the weakly labeled samples, including:
[0253] According to the strongly labeled samples and the weakly labeled samples, the strength of the characteristic relationship between the characteristic parameter data in the power equipment is obtained;
[0254] According to the strength of the characteristic relationship between the characteristic parameter data in the power equipment, the knowledge function corresponding to each set of characteristic relationships is constructed.
[0255] The expression corresponding to the cross entropy loss function of the health assessment model in the above state assessment module is as follows:
[0256]
[0257] Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ 1 represents a hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 Indicates the weight coefficient value corresponding to the first group of feature relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationships between the characteristic parameter data α and the characteristic parameter data β.
[0258] The knowledge function E corresponding to the first set of characteristic relations between the characteristic parameter data α and the characteristic parameter data β in the above-mentioned state evaluation module isl The expression of (α,β) is as follows:
[0259]
[0260] Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,n}; n represents the number of groups of feature parameter data forming the feature relations; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data at time X; Ω G represents the sampling set of characteristic parameter data at time G; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value.
[0261] In the above-mentioned state assessment module, after obtaining the strength of the characteristic relationship between each characteristic parameter data in the power equipment according to the strong labeled samples and the weak labeled samples, before constructing the knowledge function corresponding to each set of characteristic relationships according to the strength of the characteristic relationship between each characteristic parameter data in the power equipment, it also includes:
[0262] Taking each feature parameter data as a vertex, taking the feature relationship between each feature parameter data as an edge, taking the strength of the feature relationship between each feature parameter data as an edge weight, a directed weighted graph between each feature parameter data is constructed;
[0263] For each feature parameter data, according to the attention weight between the feature parameter data and its adjacent feature parameter data, the edge weight between the feature parameter data and its adjacent feature parameter data is updated; the updated edge weights are processed in descending order; according to the edge weights processed in descending order, the edges corresponding to the edge weights sorted after the preset position in the edge weights processed in descending order are deleted;
[0264] Determine the strength of the feature relationship between each feature parameter data according to the deleted edge weight corresponding to each feature parameter data;
[0265] Among them, the edge weight between the feature parameter data and its adjacent connected feature parameter data in the directed weighted graph is updated using the following expression:
[0266]
[0267] In the formula, represents the updated edge weight value of the l′th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; T i,u represents the set of messages transmitted from characteristic parameter data u in power equipment i to characteristic parameter data v of its adjacent connection at all times; represents the edge weight value of the l′-1th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; Represents the attention weight of the l′th edge between the feature parameter data u in the power equipment i and its adjacent feature parameter data v at time t.
[0268] The health status levels in the above-mentioned status assessment module include: critical defect, serious defect, general defect, sub-health, and health.
[0269] The above-mentioned status assessment module is specifically used for:
[0270] Perform high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data;
[0271] Classifying the representation vectors of the characteristic parameter data according to time to obtain the representation vectors of the characteristic parameter data at the same time and the representation vectors of the characteristic parameter data at different times;
[0272] The graph neural network is used to aggregate the representation vectors of the feature parameter data at the same time to obtain the aggregated feature vector at the same time;
[0273] The attention mechanism is used to aggregate the representation vectors of feature parameter data at different times to obtain aggregated feature vectors at different times;
[0274] Concatenate the aggregated feature vectors at the same time and the aggregated feature vectors at different times to obtain a graph representation vector;
[0275] The graph representation vector is input into the cross entropy loss function of the health assessment model of the power equipment to obtain the probability values corresponding to different health status levels of the power equipment;
[0276] The health status level of the power equipment is obtained according to the probability values corresponding to different health status levels of the power equipment.
[0277] In the above-mentioned state assessment module, after high-dimensional feature mapping is performed on the characteristic parameter data of the power equipment to obtain the representation vector of the characteristic parameter data, the representation vector of the characteristic parameter data is classified according to time to obtain the representation vector of the characteristic parameter data at the same time and the representation vector of the characteristic parameter data at different times, and further includes:
[0278] Determine whether there is missing characteristic parameter data in the power equipment, and when there is missing characteristic parameter data, obtain a representation vector of adjacent characteristic parameter data of the missing characteristic parameter data;
[0279] Calculate the attention weights between the missing feature parameter data and the adjacent feature parameter data;
[0280] According to the attention weight, the representation vector of the missing feature parameter data is obtained;
[0281] Among them, the calculation formula of the attention weight between the missing feature parameter data and the adjacent feature parameter data is as follows:
[0282]
[0283] In the formula, represents the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t; μ(·) represents the message transfer function; represents the representation vector of the characteristic parameter data u in the power equipment i at time t under the high-dimensional feature mapping; D represents the trainable weight matrix; r v A trainable weight vector representing the feature parameter data v; Represents the time feature representation vector of the feature parameter data u at time t; T represents transpose.
[0284] Embodiment 4:
[0285] The present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.
[0286] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding functions, so as to realize the steps of a data-knowledge fusion-driven power equipment status assessment method in the above-mentioned embodiment.
[0287] Embodiment 5:
[0288] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of a data-knowledge fusion-driven power equipment status assessment method in the above embodiment.
[0289] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0290] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0291] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0292] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0293] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims to be approved.
Claims
1. A data-knowledge fusion driven power equipment status assessment method, characterized in that: include: Acquire characteristic parameter data of power equipment; Based on the characteristic parameter data of the power equipment, a pre-built health assessment model of the power equipment is solved using a Raindrop learning algorithm to obtain a health status level of the power equipment; The health assessment model is constructed based on strongly labeled samples of the power equipment acquired in advance and weakly labeled samples obtained by formally expressing the knowledge and experience of the power equipment.
2. The method according to claim 1, characterized in that The health assessment model of the power equipment includes the following construction process: Obtain strong annotation samples and knowledge and experience of the power equipment; Formalize the knowledge and experience to obtain weakly labeled samples; According to the strongly labeled samples, construct a cross entropy loss function of the strongly labeled samples; According to the weakly labeled samples, construct a knowledge regularization term of the weakly labeled samples; Constructing a knowledge function according to the strongly labeled samples and the weakly labeled samples; Obtaining a cross entropy loss function of the health assessment model of the power equipment according to the cross entropy loss function of the strongly labeled samples, the knowledge regularization term of the weakly labeled samples and the knowledge function; The health assessment model of the power equipment is obtained with the goal of minimizing the cross entropy loss function of the health assessment model of the power equipment.
3. The method according to claim 2, characterized in that The constructing a knowledge function according to the strongly labeled samples and the weakly labeled samples includes: According to the strongly labeled samples and the weakly labeled samples, the strength of the characteristic relationship between the characteristic parameter data in the electric power equipment is obtained; According to the strength of the characteristic relationship between the characteristic parameter data in the power equipment, a knowledge function corresponding to each set of characteristic relationships is constructed.
4. The method according to claim 3, characterized in that The expression corresponding to the cross entropy loss function of the health assessment model is as follows: Among them, L CE (θ) represents the cross entropy loss function of the parameter θ in the health assessment model; L D (·) represents the cross entropy loss function of strongly labeled samples; y (i) represents the labeling result of the i-th power device in the strongly labeled sample; i = 1…M; M represents the number of power devices in the strongly labeled sample; h(x; θ) represents the output probability distribution when the input variable is x and the parameter is θ; λ1 represents the hyperparameter; L K (·) represents the knowledge regularization term of the weakly labeled samples; N represents the number of power equipment in the weakly labeled samples; represents the labeling result of the jth power equipment in the weakly labeled sample; j = 1…N; λ l+1 represents the weight coefficient value corresponding to the lth group of characteristic relations; l = 1…n; n represents the number of groups of characteristic parameter data forming the characteristic relations; E l (α, β) represents the knowledge function corresponding to the lth set of characteristic relationships between the characteristic parameter data α and the characteristic parameter data β.
5. The method according to claim 4, characterized in that The knowledge function E corresponding to the first set of characteristic relations between the characteristic parameter data α and the characteristic parameter data β l The expression of (α,β) is as follows: Among them, f l (·) represents the penalty function of the lth group of feature relations; l∈{1,…,n}; n represents the number of groups of feature parameter data forming the feature relations; e i,αβ represents the edge weight between the characteristic parameter data α and the characteristic parameter data β of the i-th power equipment; Ω X Represents the sampling set of characteristic parameter data X at each time; Ω G Represents the sampling set of characteristic parameter data G at each time; ε represents the strongly correlated parameter value; σ represents the weakly correlated parameter value.
6. The method according to claim 3, characterized in that After obtaining the strength of the characteristic relationship between each characteristic parameter data in the power equipment according to the strongly labeled samples and the weakly labeled samples, and before constructing the knowledge function corresponding to each group of characteristic relationships according to the strength of the characteristic relationship between each characteristic parameter data in the power equipment, the method further includes: Taking each of the feature parameter data as a vertex, taking the feature relationship between each of the feature parameter data as an edge, taking the strength of the feature relationship between each of the feature parameter data as an edge weight, and constructing a directed weighted graph between each of the feature parameter data; For each of the feature parameter data, according to the attention weight between the feature parameter data and the feature parameter data adjacent to it, the edge weight between the feature parameter data and the feature parameter data adjacent to it is updated; the updated edge weights are processed in descending order; according to the edge weights processed in descending order, the edges corresponding to the edge weights sorted after the preset position in the edge weights processed in descending order are deleted; Determining the strength of the characteristic relationship between each of the characteristic parameter data according to the deleted edge weights corresponding to each of the characteristic parameter data; The edge weight between the feature parameter data and its adjacent feature parameter data in the directed weighted graph is updated using the following expression: In the formula, The lth distance between the characteristic parameter data u in the power device i and the characteristic parameter data v connected to it ′ The updated edge weight value of the edge; T i,u represents the set of messages transmitted from characteristic parameter data u in power equipment i to characteristic parameter data v of its adjacent connection at all times; represents the edge weight value of the l′-1th edge between the characteristic parameter data u in the power device i and its adjacent characteristic parameter data v; Represents the attention weight of the l′th edge between the feature parameter data u in the power equipment i and its adjacent feature parameter data v at time t.
7. The method according to claim 1, characterized in that The health status levels include: critical defects, serious defects, general defects, sub-health, and health.
8. The method according to claim 1 or 7, characterized in that The method of solving the pre-built health assessment model of the power equipment based on the characteristic parameter data of the power equipment by using the Raindrop learning algorithm to obtain the health status level of the power equipment includes: Performing high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain a representation vector of the characteristic parameter data; Classifying the representative vectors of the characteristic parameter data according to time to obtain representative vectors of the characteristic parameter data at the same time and representative vectors of the characteristic parameter data at different times; A graph neural network is used to aggregate the representation vectors of the feature parameter data at the same time to obtain an aggregated feature vector at the same time; Aggregating the representation vectors of the feature parameter data at different times using an attention mechanism to obtain aggregated feature vectors at different times; Performing vector concatenation of the aggregated feature vectors at the same time and the aggregated feature vectors at different times to obtain a graph representation vector; Inputting the graph representation vector into the cross entropy loss function of the health assessment model of the power equipment to obtain probability values corresponding to different health status levels of the power equipment; The health status level of the power equipment is obtained according to the probability values corresponding to different health status levels of the power equipment.
9. The method according to claim 8, characterized in that After performing high-dimensional feature mapping on the characteristic parameter data of the electric power equipment to obtain the representation vector of the characteristic parameter data, and before classifying the representation vector of the characteristic parameter data according to time to obtain the representation vector of the characteristic parameter data at the same time and the representation vector of the characteristic parameter data at different times, the method further includes: Determine whether there is missing characteristic parameter data in the power equipment, and when there is missing characteristic parameter data, obtain a representation vector of adjacent characteristic parameter data of the missing characteristic parameter data; Calculating the attention weight between the missing feature parameter data and the adjacent feature parameter data; According to the attention weight, obtaining a representation vector of the missing feature parameter data; The calculation formula of the attention weight between the missing feature parameter data and the adjacent feature parameter data is as follows: In the formula, represents the attention weight between the characteristic parameter data u and the characteristic parameter data v in the power equipment i at time t; μ(·) represents the message transfer function; represents the representation vector of the characteristic parameter data u in the power equipment i at time t under the high-dimensional feature mapping; D represents the trainable weight matrix; r v A trainable weight vector representing the feature parameter data v; Represents the time feature representation vector of the feature parameter data u at time t; T represents transpose.
10. A data-knowledge fusion driven power equipment status assessment system, characterized in that: include: Data acquisition module: used to obtain characteristic parameter data of power equipment; A status assessment module is used to solve the pre-built health assessment model of the power equipment based on the characteristic parameter data of the power equipment using the Raindrop learning algorithm to obtain the health status level of the power equipment; The health assessment model in the state assessment module is constructed based on the strongly labeled samples of the power equipment acquired in advance and the weakly labeled samples obtained by formally expressing the knowledge and experience of the power equipment.
