Fan fault decision-making method based on dynamic evolution of area, model and knowledge graph

Abstract

The invention relates to the technical field of operation and maintenance of new energy electric field fan equipment, in particular to a fan fault decision-making method based on region, model and knowledge graph dynamic evolution, which comprises the following steps: S1, constructing a maintenance experience database according to historical data, and updating data according to a newly added maintenance work order; s2, carrying out quantitative modeling according to regional features and model features in the maintenance experience database, constructing a feature matrix, and carrying out dynamic updating according to a newly added maintenance work order; s3, constructing a knowledge graph according to the maintenance experience database, fusing the feature matrix and the knowledge graph, and performing dynamic knowledge graph updating through real-time data driving and closed-loop feedback optimization double-introduction; and S4, adopting a self-optimization aided decision, predicting a fault occurrence probability through a reinforcement learning strategy network, and matching a solution in the knowledge graph. The method is suitable for wind generating set fault auxiliary decision making, and the wind power operation management efficiency is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of operation, maintenance and repair of new energy power plant fan equipment, and specifically relates to a fan fault decision-making method based on regional, model and dynamic evolution of knowledge graph. Background Art

[0002] Wind turbines are core equipment in the new energy field, and their normal operation is crucial for ensuring energy security, reducing accident risks and avoiding economic losses. However, the working environment of wind turbines is harsh and the operating states change frequently, resulting in frequent failures of their key components. There are often more than a dozen fault troubleshooting and repair methods for the same scada fault name. Traditional repair methods rely on the experience judgment of engineers, which not only takes time and effort, but is also easily interfered by human factors, leading to low repair efficiency and even possible delay in the best time for fault handling.

[0003] Existing artificial intelligence algorithm models either rely on static data cleaning and decision tree models, lacking dynamic knowledge update and multi-dimensional correlation analysis capabilities; or use static knowledge graphs and fixed inference rules, unable to reflect changes in fault patterns in real time, and lacking a feedback optimization mechanism for decision recommendations. At the same time, these models do not consider the influence of fan model differences (such as direct drive type, doubly fed type) and regional environmental characteristics (such as high salt fog in coastal areas, wind sand in inland areas) on fault patterns, resulting in a decline in the performance of general models in specific scenarios. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a fan fault decision-making method based on regional, model and dynamic evolution of knowledge graph, which is applicable to fault auxiliary decision-making of wind turbines and greatly improves the efficiency of wind power operation management.

[0005] The technical solution adopted by the present invention is to provide a fan fault decision-making method based on regional, model and dynamic evolution of knowledge graph, including the following steps,

[0006] S1. Construct a maintenance experience database based on historical data and update the data according to new maintenance work orders. The maintenance experience database includes region, climate, terrain data, region-specific fault patterns, power stations, fan parameters, faulty components, fault codes, fault phenomena, fault causes and solutions;

[0007] S2. Quantify and model according to the regional characteristics and model characteristics in the maintenance experience database, construct a feature matrix, and dynamically update it according to new maintenance work orders;

[0008] S3. Construct a knowledge graph based on the maintenance experience database, fuse the feature matrix with the knowledge graph, and perform dynamic knowledge graph update through dual guidance of real-time data driving and closed-loop feedback optimization;

[0009] S4. Adopt self-optimizing auxiliary decision-making, predict the probability of failure through the reinforcement learning policy network, sort them in descending order according to the probability of failure, and match the solutions in the knowledge graph.

[0010] The specific steps of step S1 include:

[0011] S1.1. Initialize historical data,

[0012] Clean and extract experience from the existing standard data of the fan and historical maintenance data. Filter out incorrect, incomplete, and incorrectly formatted data through a recursive traversal method, and then extract it after data identification, understanding, screening, and induction. Form structured historical maintenance experience data and save it in the maintenance experience database.

[0013] S1.2. Update the maintenance experience database,

[0014] After a new maintenance work order is executed, automatically extract the relevant data in the work order and enter it into the maintenance experience database.

[0015] The specific steps of step S2 include:

[0016] S2.1. Quantitatively calculate the salt spray corrosion coefficient C1, and the formula is as follows:

[0017]

[0018] In the formula, S avg represents the annual average salt spray concentration in the area, S rc represents the reference concentration in the area, R mc represents the material corrosion rate coefficient;

[0019] S2.2. Quantitatively calculate the wind and sand abrasion coefficient C2, and the formula is as follows:

[0020]

[0021] In the formula, W avg represents the annual average wind speed in the area, W t represents the wind resistance limit of the fan, S d represents the dust density, W c represents the blade abrasion resistance coefficient;

[0022] S2.3. Quantitatively calculate the altitude heat dissipation correction factor C3, and the formula is as follows:

[0023]

[0024] In the formula, H a represents the altitude;

[0025] S2.4. Quantitative calculation of the terrain vibration coefficient C4, with the formula as follows:

[0026]

[0027] In the formula, 1.2 is the terrain vibration coefficient for mountains and rock formations, 1.0 is the vibration coefficient for sandy land, and 0.8 is the terrain vibration coefficient for plains and clay lands.

[0028] S2.5. Regional historical fault weight feature C si Quantitative calculation, with the formula as follows:

[0029]

[0030] In the formula, N i represents the occurrence times of fault i in the region, N r represents the total number of faults in the region, and F ic represents the influence coefficient of fault i.

[0031] S2.6. Quantitative calculation of the model feature C6, with the formula as follows:

[0032] C6 = 0.7×FP0 + 0.3×FP a

[0033] In the formula, FP0 represents the design failure rate, and FP a represents the failure rate of this model in the region.

[0034] S2.7. Fuse the quantified regional features.

[0035] CV = [C1, C2, C3, C4, C 51 , C 52 ,..., C 5i , C6]

[0036] In the formula, CV represents the fused feature vector.

[0037] S2.8. Dynamically update the feature matrix.

[0038] After the commissioning of new fan models and changes in regional environmental data, when the feedback of the maintenance work order indicates that the matrix prediction accuracy rate is lower than 25%, the feature matrix is automatically updated dynamically.

[0039] The specific steps of S3 are as follows:

[0040] S3.1. Construct a knowledge graph based on the maintenance experience database, and inject each feature vector in the feature matrix as an attribute into the knowledge graph nodes.

[0041] S3.2. Dynamically generate the ternary relationship of region, model, and fault, representing the occurrence probability of faults under each combination of region and model. The formula is as follows:

[0042] W tr = ω1×C1 + ω2×C2 + ω3×C3 + ω4×C4 + ω5×C 5i + ω6×C6

[0043] In the formula, ω1, ω2, ω3, ω4, ω5, and ω6 respectively represent the characteristic weights of C1, C2, C3, C4, C 5i , and C6;

[0044] S3.3. Dynamically update the knowledge graph through real-time data-driven and closed-loop feedback optimization of the dual engines. The update formula is as follows:

[0045]

[0046] In the formula, η represents the learning rate, with a default value of 0.1, and M d represents the matching degree feedback in the work order, ranging from 0 to 1.

[0047] The specific steps of step 4 include:

[0048] S4.1. According to the fault code, fan area, and fan model information received in real time, obtain the fault association relationship and weight under the current area and model combination from the knowledge graph;

[0049] S4.2. Integrate real-time data, area, model, and knowledge graph weights to calculate the reinforcement learning state vector S v , and the formula is as follows:

[0050] S v = [F c , C1, C2, C3, C4, C6, W n , ΔW n

[0051] In the formula, F c represents the fault code, C1, C2, C3, C4, C6 represent each feature vector, W n represents the fault weight under the area and model combination, and ΔW n represents the change trend of the fault weight under the area and model combination;

[0052] S4.3. Input the vector S v from the previous step into the trained reinforcement learning policy network, and output the fault distribution with the highest occurrence probability. The formula is as follows:

[0053] P sv = Softmax(W3·ReLU(W2·ReLu(W1·S v + b1)+ b2)+ b3)

[0054] ​Wherein, W1, W2, and W3 are respectively the randomly initialized weight matrices of hidden layers 1-3, b1, b2, and b3 are respectively the randomly initialized bias vectors of hidden layers 1-3, ReLU is an activation function, and Softmax is an activation function;

[0055] S4.3. Match the solution from the knowledge graph according to the fault with the highest occurrence probability in the previous step, and give a recommendation value.

[0056] The beneficial effects of the present invention are as follows: A fan fault decision-making method based on regional, model, and dynamic evolution of the knowledge graph is provided. Aiming at the problem of lack of maintenance auxiliary decision-making after the current wind turbine generator set fails to alarm, it aims to construct a multi-dimensional dynamic knowledge graph integrating regional environment, fan model, and fault entities, and combine a differential reinforcement learning algorithm with a cross-domain transfer learning mechanism to achieve precise and adaptive optimization of fault diagnosis. Aiming at the problems of static knowledge base and poor universality of decision-making models in the prior art, the present invention introduces regional and model feature matrices to quantify the influence of environmental and equipment differences on fault modes, and designs dynamic knowledge evolution to solve the data cold start problem in new model / new region scenarios. By real-time updating the graph weights and decision-making strategies through closed-loop feedback, in the actual measurements of multiple wind farms, the fault diagnosis time is significantly shortened, the false alarm rate in cross-model scenarios is significantly reduced, the maintenance cost is effectively reduced, and the intelligent level and economic benefits of fan fault decision-making are significantly improved. Description of the Drawings

[0057] Figure 1 is the flow block diagram of the present invention. Detailed Embodiments

[0058] The embodiments of the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] As Figure 1 shown, the present invention provides a fan fault decision-making method based on regional, model, and dynamic evolution of the knowledge graph, including the following steps:

[0060] S1. Construction of a maintenance experience database. Clean and transform the existing fan standard data, feedback information from on-site maintenance personnel during the work order execution process, and historical maintenance data to construct a modular and hierarchical fault experience database.

[0061] S1.1. Initialization of historical data

[0062] The existing fan standard data and historical maintenance data are subjected to data cleaning and experience extraction. After filtering out incorrect, incomplete, and incorrectly formatted data through a recursive traversal method, they are then extracted after data recognition, understanding, screening, and induction to form structured historical maintenance experience data, which is saved in the maintenance experience database.

[0063] The maintenance experience database includes region, climate, terrain data, region-specific fault modes, power stations, fan parameters, faulty components, fault codes, fault phenomena, fault causes, and solutions, which are saved in the maintenance experience database as this fault record. Among them, climate includes temperature and humidity, salt fog concentration, and wind speed distribution; terrain data includes altitude, surface type, and grid access stability; fan parameters include model, direct drive / dual-fed type, gearbox material, and blade length.

[0064] S1.2. Update of the experience database

[0065] After the maintenance work order is executed, relevant fields in the fault work order are extracted according to the registered work order feedback information and saved in the maintenance experience database.

[0066] S2. Quantify and model the region and model characteristics. Divide the region characteristics into three categories: climate attributes, terrain attributes, and historical fault attributes, achieve precise quantification of relevant characteristics, and provide a differential decision-making basis for subsequent fault diagnosis through dynamic coupling with model characteristics.

[0067] S2.1. Quantitatively calculate the salt fog corrosion coefficient C1. The formula is as follows

[0068]

[0069] In the formula, S avg represents the annual average salt fog concentration in the region, S rc represents the reference concentration in the region, R mc represents the material corrosion rate coefficient;

[0070] S2.2. Quantitatively calculate the wind and sand abrasion coefficient C2. The formula is as follows

[0071]

[0072] In the formula, W avg represents the annual average wind speed in the region, W t represents the wind resistance limit of the fan, S d represents the dust density, W c represents the blade abrasion resistance coefficient;

[0073] S2.3. Quantitatively calculate the altitude heat dissipation correction factor C3. The formula is as follows

[0074]

[0075] In the formula, H a represents the altitude;

[0076] S2.4. Quantitatively calculate the terrain vibration coefficient C4. The formula is as follows

[0077]

[0078] In the formula, 1.2 is the terrain vibration coefficient for mountains and rock formations, 1.0 is the vibration coefficient for sandy land, and 0.8 is the vibration coefficient for plains and clay terrains;

[0079] S2.5. Quantitatively calculate the regional historical fault weight feature C si The formula is as follows

[0080]

[0081] In the formula, N i represents the occurrence times of fault i in the region, N r represents the total number of faults in the region, and F ic represents the influence coefficient of fault i. The influence coefficient is divided into 5 levels: minor = 1, first level = 2, second level = 3, third level = 4, major = 5;

[0082] S2.6. Quantitatively calculate the model feature C6. The formula is as follows

[0083] C6 = 0.7×FP0 + 0.3×FP a

[0084] In the formula, FP0 represents the design failure rate, and FP a represents the failure rate of this model in the region;

[0085] S2.7. Integrate the quantified regional features

[0086] CV = [C1, C2, C3, C4, C 51 , C 52 ,..., C 5i , C6]

[0087] In the formula, CV represents the integrated feature vector;

[0088] S2.8. Dynamically update the feature matrix. After a new fan model is put into operation and the regional environmental data changes, when the feedback of the maintenance work order shows that the prediction accuracy of the matrix is lower than 25%, the feature matrix is automatically updated dynamically.

[0089] S3. Dynamically fuse the knowledge graph, automatically adapt the result of the feature matrix in the previous step to the knowledge graph, without manual reconstruction of rules.

[0090] S3.1. Based on the knowledge graph constructed from the maintenance experience database, inject the region feature vector and model feature vector generated in step 2 as attributes into the knowledge graph nodes to expand the original attributes.

[0091] S3.2. Dynamically generate the three - element relationship of region, model, and fault, which characterizes the occurrence probability of faults under a specific region - model combination. The formula is as follows:

[0092] W tr = ω1×C1 + ω2×C2 + ω3×C3 + ω4×C4 + ω5×C 5i + ω6×C6

[0093] In the formula, ω1, ω2, ω3, ω4, ω5, ω6 respectively represent the feature weights of C1, C2, C3, C4, C 5i , C6.

[0094] S3.3. Support the update of the dynamic knowledge graph through the dual - engine of real - time data - driven and closed - loop feedback optimization to ensure the timeliness and accuracy of the graph relationship weights. The real - time data - driven is mainly updated automatically after the dynamic update of the feature matrix, and the closed - loop feedback optimization dual - engine is updated through the work order feedback results. The update formula is as follows:

[0095]

[0096] In the formula, η represents the learning rate, with a default value of 0.1, and M d represents the matching degree feedback in the work order, ranging from 0 to 1;

[0097] S4. Self - optimization assisted decision - making, automatically recommend the optimal fault - solving solution to engineers through the self - optimization engine algorithm.

[0098] S4.1. According to the fault code F c , fan region, and fan model information received in real - time, obtain the fault association relationship and weight W n under the current region - model combination from the knowledge graph.

[0099] S4.2. Integrate real - time data, region - model encoding, and graph weights to construct a reinforcement learning state vector. The formula is as follows:

[0100] S v =[F c , C1, C2, C3, C4, C6, W n , ΔW n

[0101] In the formula, F c represents the fault code, C1, C2, C3, C4, C6 represent each feature vector, and W n ​Represents the fault weight under the combination of region and model, ΔW n Represents the changing trend of the fault weight under the combination of region and model.

[0102] S4.3. Input the vector result of the previous step into the trained reinforcement learning policy network, and output the fault distribution with the highest probability. The reinforcement learning policy network is a three-layer fully connected network, the activation function is ReLU, and the forward propagation formula used by the policy network is as follows:

[0103] P sv = Softmax(W3·ReLU(W2·ReLu(W1·S v + b1)+ b2)+ b3)

[0104] In the formula, W1, W2, and W3 are the randomly initialized weight matrices of hidden layers 1 - 3 respectively, b1, b2, and b3 are the randomly initialized bias vectors of hidden layers 1 - 3 respectively, ReLU is the activation function, and Softmax is the activation function.

[0105] S4.4. According to the fault with the highest probability in the previous step, match the solution from the knowledge graph and give the recommendation value information to facilitate the user to select the solution.

[0106] The present invention realizes the precision and adaptive optimization of fan fault diagnosis by constructing a multi-dimensional dynamic knowledge graph integrating regional environment, fan model, and fault entities, and combining a differential reinforcement learning algorithm. Compared with traditional methods, this solution significantly shortens the fault diagnosis time, greatly reduces the false alarm rate in cross-model scenarios, effectively solves the problems of knowledge update and feedback optimization, and the cold start problem during the deployment of new models and new regions. Through accurate decision-making recommendations and a dynamic knowledge evolution mechanism, the maintenance cost is greatly reduced, and the economic benefits and intelligent level of wind farm operation and maintenance are improved.

[0107] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A wind turbine fault decision-making method based on the dynamic evolution of region, model and knowledge graph, characterized by: The following steps are included: S1. Build a maintenance experience database based on historical data and update the data based on newly added maintenance work orders. The maintenance experience database includes regional, climate, terrain data, regional specific failure modes, stations, wind turbine parameters, faulty components, fault codes, fault phenomena, fault causes and solutions; S2. Quantitative modeling is performed based on the regional characteristics and model characteristics in the maintenance experience database, a feature matrix is ​​constructed, and dynamic updates are performed based on newly added maintenance work orders; S3. Build a knowledge graph based on the maintenance experience database, merge the feature matrix with the knowledge graph, and dynamically update the knowledge graph through real-time data drive and closed-loop feedback optimization. S4. Use self-optimization to assist decision making, predict the probability of fault occurrence through reinforcement learning strategy network, sort them from high to low according to the probability of fault occurrence, and match the solution in the knowledge graph.

2. The wind turbine fault decision-making method based on region, model and dynamic evolution of knowledge graph according to claim 1 is characterized in that: The step S1 specifically includes: S1.1, historical data initialization, The existing fan standard data and historical maintenance data are cleaned and experience extracted. The wrong, incomplete and incorrectly formatted data are filtered out through recursive traversal. After data identification, understanding, screening and induction, the data are extracted to form structured historical maintenance experience data and save them in the maintenance experience database. S1.2, Maintenance experience database update, After the newly added maintenance work order is executed, the relevant data in the work order is automatically extracted and entered into the maintenance experience database.

3. The wind turbine fault decision-making method based on region, model and dynamic evolution of knowledge graph according to claim 1 is characterized in that: The step S2 specifically includes: S2.

1. Quantitative calculation of salt spray corrosion coefficient C1, the formula is as follows: In the formula, S avg represents the average annual salt spray concentration in the region, S rc Indicates the regional reference concentration, R mc Indicates the material corrosion rate coefficient; S2.

2. Quantitative calculation of wind and sand wear coefficient C2, the formula is as follows: Where W avg represents the average annual wind speed in the region, W t Indicates the wind resistance limit of the fan, S d represents dust density, W c Indicates the blade wear resistance coefficient; S2.

3. Quantitative calculation of altitude heat dissipation correction factor C3 is shown as follows: In the formula, H a Indicates altitude; S2.4, Quantitative calculation of terrain vibration coefficient C4, the formula is as follows: In the formula, 1.2 is the vibration coefficient of mountainous and rocky terrain, 1.0 is the vibration coefficient of sandy terrain, and 0.8 is the vibration coefficient of plain and clay terrain; S2.

5. Regional historical fault weight characteristics C si The quantitative calculation formula is as follows: Where N i represents the number of occurrences of fault i in area, N r represents the total number of faults in the region, F ic represents the influence coefficient of fault i; S2.6, model feature C6 quantitative calculation, the formula is as follows, C6=0.7×FP0+0.3×FP a Where FP0 represents the design failure rate, FP a Indicates the failure rate of this model in the area; S2.7, fuse the quantified regional features, CV=[C1,C2,C3,C4,C 51 ,C 52 ,...,C 5i ,C6] In the formula, CV represents the fused feature vector; S2.8, dynamic update of feature matrix, When new wind turbine models are put into operation and regional environmental data change, and maintenance work order feedback indicates that the matrix prediction accuracy is less than 25%, the feature matrix is ​​automatically updated dynamically.

4. The wind turbine fault decision method based on region, model and dynamic evolution of knowledge graph according to claim 3 is characterized by: The step S3 specifically includes: S3.

1. Build a knowledge graph based on the maintenance experience database, and inject each feature vector in the feature matrix as an attribute into the knowledge graph node; S3.2, dynamically generate the ternary relationship between region, model, and fault to represent the probability of fault occurrence under each region and model combination. The formula is as follows: IN tr =ω1×C1+ω2×C2+ω3×C3+ω4×C4+ω5×C 5i +ω6×C6 In the formula, ω1, ω2, ω3, ω4, ω5, and ω6 represent C1, C2, C3, C4, and C 5i , feature weight of C6; S3.

3. Dynamically update the knowledge graph through the dual engines of real-time data drive and closed-loop feedback optimization. The update formula is as follows: In the formula, η represents the learning rate, the default value is 0.1, M d Indicates the matching degree of the feedback in the work order, which is 0-1.

5. The wind turbine fault decision method based on region, model and dynamic evolution of knowledge graph according to claim 4 is characterized in that: The step 4 specifically includes: S4.

1. According to the fault code, wind turbine region and wind turbine model information received in real time, the fault association relationship and weight under the current region and model combination are obtained from the knowledge graph; S4.

2. Integrate real-time data, region, model, and knowledge graph weights to calculate the reinforcement learning state vector S v , the formula is as follows, S v =[F c ,C1,C2,C3,C4,C6,W n ,ΔW n ] In the formula, F c represents the fault code, C1, C2, C3, C4, C6 represent the feature vectors, W n represents the fault weight under the region and model combination, ΔW n Indicates the changing trend of fault weights under region and model combination; S4.

3. Input the previous step vector S v To the trained reinforcement learning strategy network, output the fault distribution with the highest probability value. The formula is as follows: P sv =Softmax(W3·ReLU(W2·ReLu(W1·S v +b1)+b2)+b3) where W1, W2, W3 are randomly initialized weight matrices of hidden layers 1-3, b1, b2, b3 are randomly initialized bias vectors of hidden layers 1-3, ReLU is the activation function, and Softmax is the activation function; S4.

3. According to the fault with the highest probability of occurrence in the previous step, match the solution from the knowledge graph and give a recommendation value.

Images