Wind power online fault diagnosis system based on knowledge graph

Abstract

The invention discloses a wind turbine generator line autonomous learning diagnosis system based on a knowledge graph. The method comprises the following steps: firstly, extracting fault feature parameters of fan state monitoring data by using a Faster RCNN model, and then converging fault features to obtain a fault diagnosis standard; secondly, constructing a fan intelligent operation and maintenance knowledge graph through a state monitoring knowledge base, and combining a fault diagnosis module with the knowledge graph to construct a fan automatic diagnosis system; and finally, comparing a fan fault diagnosis result with reality, feeding back an actual result to a state detection knowledge base, carrying out iterative updating on the state detection knowledge base, feeding back a fault degree to a diagnosis standard generation module, and carrying out iterative updating on a diagnosis standard to complete autonomous learning and iterative updating of the fan.

Description

Technical Field

[0001] The invention belongs to the technical field of wind power generation and relates to a method for online status monitoring of a wind turbine. Background Art

[0002] As the global fossil energy crisis becomes increasingly serious, the energy needs of various countries are also changing. Traditional, non-renewable energy that was originally based on fossil energy is now slowly transitioning to clean and renewable energy represented by wind energy and solar energy.

[0003] In recent years, wind energy, as a typical representative of clean energy, has attracted global attention and attention, which has provided a good opportunity for the development of wind power generation.

[0004] Wind turbines usually operate in remote areas with harsh environments and are often affected by severe weather. In addition, as the capacity of single wind turbines increases and the machine structure becomes increasingly complex, these conditions determine that wind turbines have the characteristics of high failure probability and multiple types of failures, making wind turbine operation and maintenance more complicated.

[0005] At present, traditional wind power operation and maintenance is mainly carried out through manual inspections supplemented by regular maintenance. However, when abnormalities occur, operation and maintenance personnel often cannot get timely warnings. In addition, this type of operation and maintenance has problems such as low efficiency, high cost, long downtime, and high risk factor, which directly affects the safety and economy of wind power.

[0006] In terms of fault detection technology, the wind turbine condition monitoring system (CMS) system has been widely used in wind farm operation and maintenance, and effectively provides early warning and alarm for wind turbine status.

[0007] However, CMS still has some problems. First, CMS data often focuses on detecting and collecting status data of wind turbines, and cannot characterize the overall structure of the unit and update and infer faults from its abnormal data; second, CMS cannot accurately and effectively identify some low-frequency faults in real time and provide timely warnings; finally, the current CMS system can only evaluate the operating status of corresponding components based on a known knowledge base, and cannot automatically adjust to situations outside the knowledge base. Summary of the Invention

[0008] This method combines condition monitoring data with the wind turbine fault diagnosis knowledge base, constructs an online automatic diagnosis system based on the operation and maintenance Internet of Things and knowledge graph, combines the fault knowledge graph with wind turbine diagnosis to perform real-time data monitoring, can characterize the overall monitoring structure of the wind turbine, and effectively identify the characteristics of some faults and provide timely warnings; and proposes a wind turbine generator condition monitoring method with autonomous learning ability, which improves the accuracy and adaptability of wind turbine condition monitoring, infers fault characteristics, and provides new ideas for wind farm operation and maintenance personnel to warn, locate and solve faults.

[0009] like Figure 1 As shown in Figure 2, this method consists of four layers: data acquisition layer, knowledge acquisition layer, knowledge expression layer, and knowledge learning layer. The data acquisition layer is responsible for collecting and transmitting wind turbine vibration monitoring data. The condition monitoring data includes the vibration acceleration data of the wind turbine transmission chain, blades, tower, and nacelle, such as Figure 2 As shown in the figure, the collected condition monitoring data is transmitted to the wind turbine's local intelligent terminal for integration, naming, and time-frequency transformation, preparing the knowledge acquisition layer for extracting fault sensitivity factors for each wind turbine component. The knowledge acquisition layer is responsible for preparing the prior knowledge that constitutes the knowledge graph and extracting the fault sensitivity factors for each wind turbine component. This prior knowledge includes a knowledge base comprised of the experience of on-site operation and maintenance personnel and experts. The Faster RCNN module extracts features of the fault sensitivity factors from the data obtained by the data acquisition layer and aggregates them to the diagnostic criteria generation module through the operation and maintenance Internet of Things. The criteria generation module is responsible for integrating the sensitivity factors of the corresponding wind turbine components into wind farm-level features to generate diagnostic criteria for each wind turbine component. The knowledge representation layer is responsible for constructing a knowledge graph for wind turbine condition detection and building a visual condition detection system. This layer combines the condition monitoring knowledge base with the data to construct a knowledge graph that identifies the relationships between wind turbine components, fault types, operating status, maintenance recommendations, and fault causes. It monitors real-time wind turbine data, evaluates the wind turbine's operating status, and presents it through the condition monitoring system.

[0010] The knowledge learning layer is responsible for iteratively updating and autonomously learning diagnostic standards and wind turbine diagnostic experience. It evaluates the unit's operating status based on the standards obtained by the knowledge acquisition layer, compares on-site maintenance feedback with the condition monitoring standards, and feeds back inaccurate diagnostic results to the fault diagnosis experience database for knowledge learning and updating. The fault diagnosis accuracy is then calculated. If the accuracy rate does not meet the requirements, standard learning and updating are performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Online monitoring and diagnosis system for wind turbines based on knowledge graph.

[0012] Figure 2 Schematic diagram of wind turbine data collection.

[0013] Figure 3 Automatic spectrum identification process.

[0014] Figure 4 Spectral feature extraction network model.

[0015] Figure 5 Image feature extraction module.

[0016] Figure 6 RPN module.

[0017] Figure 7 Schematic diagram of the generator fault tree. DETAILED DESCRIPTION

[0018] The data acquisition layer includes data collection and transmission. Figure 2 As shown in the figure, data on the wind turbine drive chain, blades, nacelle, and tower are collected and transmitted to the local intelligent terminal of each wind turbine for time-frequency conversion. The formula for time-frequency conversion is as follows:

[0019]

[0020] Where k = 0, 1, ..., N-1;

[0021]

[0022] The output signal's spectrogram is 640x640 in size and in BGR format.

[0023] The data from each wind turbine’s local intelligent terminal is then transmitted to the knowledge acquisition layer.

[0024] The knowledge acquisition layer is used for feature extraction and collection of prior knowledge.

[0025] The local intelligent terminal of the wind turbine performs feature extraction on the data collected from each unit, and the extracted content is as follows: Figure 3 As shown, the process is as follows.

[0026] S1: The status analysis staff calculates the fault characteristic frequency of the unit based on the parameters of each component of the unit and marks it in the spectrum. The marked frequencies include:

[0027] Blade: first-order dancing frequency, first-order shimmying frequency, second-order dancing frequency, second-order shimmying frequency, blade rotation frequency, first-order torsional vibration frequency, second-order torsional vibration frequency;

[0028] Tower: first-order resonant frequency of the tower, second-order resonant frequency of the tower, first-order torsional vibration frequency, second-order torsional vibration frequency;

[0029] Spindle: spindle rotation frequency, spindle bearing fault characteristic frequency;

[0030] Gearbox: gear fault characteristic frequency of each gear in the gearbox, meshing frequency of each gear in the gearbox and its multiples, gearbox bearing fault characteristic frequency and its multiples, and gearbox support structure;

[0031] Generator: generator rotation frequency and its multiples, generator bearing fault characteristic frequency, generator pole passing frequency, power frequency, etc.

[0032] S2: Build a feature extraction network model and input the spectrum graph into the network. The network model is as follows: Figure 4 As shown:

[0033] S2.1: First, input the image into the feature extraction network to obtain the corresponding feature map.

[0034] Feature extraction module such as Figure 5 As shown in the figure, "m×n conv, X" is a convolutional layer, where the width and height of the convolution kernel are m and n respectively, and the number of channels of the convolution is X. Taking "3×3conv, 64" as an example, it means that the width and height of the convolution kernel are both 3, and the number of channels is 64.

[0035] pool l / 2 indicates the maximum pooling layer with a pooling kernel of 2*2.

[0036] S2.2: The extracted feature map is passed through the RPN (Region Proposal Network) network to generate candidate boxes. The candidate boxes generated by the RPN network are projected onto the feature map to obtain the corresponding feature matrix.

[0037] The PRN network structure is as follows Figure 4 As shown:

[0038] For each 3×3 sliding window of the feature map, we first calculate the position of the center point on the original image, and then generate a series of anchor boxes with this position as the center.

[0039] K represents the number of anchor boxes, and 2k is the two probabilities generated for each anchor point, one is the probability of being the background, and the other is the probability of being the target feature frequency. If the object in an anchor box is the feature frequency to be detected, then it is the target feature frequency, otherwise it is the background.

[0040] For each anchor box, four bounding box regression parameters are generated. The regression parameters are also a predicted offset for the center coordinates of the anchor box and an adjustment for the width and height.

[0041] Through 3×3 convolution, set the step size to 1, generate a height, width and depth with the original feature Figure 1 The characteristic matrix.

[0042] S2.3: These generated feature matrices are scaled to a uniform size through the RoI Pooling layer, ultimately to a 7×7 size. The feature matrices are then flattened and passed through a fully connected layer to obtain the predicted probabilities and bounding box regression parameters. Two 1×1 fully connected layers are connected in parallel to the resulting feature matrices to predict the bounding box regression parameters and category.

[0043] S2.4: The loss function of the RPN part consists of two parts, one is the classification loss and the other is the bounding box regression loss.

[0044]

[0045] in,

[0046] where p i Represents the probability that the i-th anchor box is predicted to be the true label;

[0047] Indicates that the current calculated sample is 1 when it is a positive sample and 0 when it is a negative sample;

[0048] t i represents the bounding box regression parameters for predicting the i-th anchor box;

[0049] Represents the regression parameters of the bounding box in the real box corresponding to the i-th anchor box;

[0050] N cls Indicates the number of samples, N in the present invention cls =256;

[0051] N reg Indicates the number of anchor box positions. In this invention, N reg =2400;

[0052] L cls is the classification loss, calculated as follows:

[0053]

[0054] L reg is the bounding box regression loss, calculated as follows:

[0055]

[0056] t i =[t x ,t y , t w , t h ]

[0057]

[0058] Calculate the loss function of the regression boundary to minimize it and obtain the final target image I. After model calculation, the eigenvalue of the target image is:

[0059] r←Faster RCNN(I)

[0060] S2.5: After feature extraction, the eigenvalues ​​are estimated. The estimation process is as follows:

[0061] F1: Corrected eigenvalue r * :r * =r+w

[0062] In the formula

[0063] w is the current noise and obeys Gaussian white noise;

[0064] F2: Define the corrected observation equation C: C * =r * +v *

[0065] In the formula

[0066] v * is the current noise, and obeys Gaussian white noise;

[0067] F3: Correct the correlation parameter P of the current eigenvalue * :P * =P+Q

[0068] In the formula

[0069] P is the updated value of the current feature value;

[0070] Q is the covariance of noise w;

[0071] F4: Update eigenvalue gain M * :M * =P * (P * +J) -1

[0072] In the formula

[0073] J is the noise v * covariance of

[0074] (P * +J) -1 For (P * +J);

[0075] F5: Estimated eigenvalues In the formula

[0076] is the estimated value of the target image.

[0077] S2.6: Calculate the loss function of Faster RCNN:

[0078] L(p,u,t u ,v)=L cls (p,u)+λ[u≥1]L loc (t u ,v)

[0079] Where P represents the softmax probability distribution predicted by the classifier P = (P0, ..., P K )

[0080] P0 represents the probability that the candidate region is the background

[0081] u represents the corresponding target true category label

[0082] t u The regression parameters of the corresponding category u predicted by the corresponding bounding box regressor

[0083] v corresponds to the bounding box regression parameters of the true target (v x ,v y ,v w ,v h )

[0084] L cls (p,u)=-logp u , is the classification loss

[0085] Among them, p u Indicates the probability that the classifier predicts that the current candidate region is category u

[0086]

[0087] S3: Input the real-time data of the wind turbine into the trained model to identify the target frequency and frequency size in the model.

[0088] The target characteristic frequency is the sensitivity factor of the data measurement point under the same working conditions, which is input into the operation and maintenance Internet of Things.

[0089] Through the operation and maintenance Internet of Things, the sensitive factors of each measuring point under the same working conditions are selected and input into the diagnostic standard generation module to first analyze the characteristic values. The algorithm steps are as follows:

[0090] T1: Determine the number of clusters to be 2;

[0091] T2: Randomly select 2 data points from the data set as the initial centroid, denoted as T3: For each sample point a i, assign it to the nearest centroid;

[0092] T4: For each class center, recalculate the center of the class:

[0093]

[0094] in

[0095] a i is the i-th sample;

[0096] c i for a i The cluster to which it belongs;

[0097] is the center of the cluster, i.e. the centroid;

[0098] M is the total number of samples.

[0099] T5: Repeat steps T3 and T4. When the centroid no longer changes, stop and output the clustering results.

[0100] The class with a large number of cases in the judgment class determines that the standard state of the component is normal, otherwise it is abnormal, where the threshold of the normal state is δ.

[0101] Construct the fault tree of the monitoring component based on the condition monitoring knowledge base and existing data. Take the generator as an example, Figure 3 As shown, and calculate the probability of each event occurring.

[0102] According to the fault tree, a knowledge graph in the form of triples of G = (E, R, S) is constructed, where E = {e1, e2, e3, ...e n} is the collection of views in the knowledge graph, R = {r1, r2, r3, ..., r n} is the set of relations in the knowledge graph. S = E × R × E represents a triple consisting of entities and relations. The basic form can be entity 1-relationship-entity 2, such as Figure 7 As shown, the generator fault includes a rotor fault.

[0103] The algorithm for knowledge graph reasoning is as follows:

[0104] E1: Consider the knowledge graph as a Markov process with a stationary distribution of π(h), inputting the state transition matrix Q, the number of state transitions n1, the number of samples n2, and the adjacency matrix x t ;

[0105] E2: Sample the initial state value h0 from any simple probability distribution;

[0106] E3: For t = 0 to n1 + n2 - 1,

[0107] E3.1: From the conditional probability distribution Q(h|h t ) Sampling to obtain sample h * ;

[0108] E3.2: Sample U ∼ [0, 1] from a uniform distribution;

[0109] E3.3: If Then accept h t+1 =h * ;

[0110] In the formula, α(h n ,h * )=π(h * )Q(h * ,h n );

[0111] E3.4: Otherwise, the transfer is not accepted. t+1 =h t ;

[0112] E4: Return sample set That is the corresponding sample set that conforms to the stationary distribution.

[0113] When an indicator in the status monitoring exceeds the alarm threshold and triggers an alarm, the on-site operation and maintenance personnel can check the fault location based on the alarm system and investigate the faulty components and the extent of the fault.

[0114] After investigation, the operation and maintenance personnel will enter the investigation results into the system.

[0115] If the faulty component does not correspond, the state monitoring experience database is updated, the knowledge graph structure is updated, and the reasoning algorithm of the above knowledge graph is retrained to calculate the fault transfer situation.

[0116] If the fault degree does not correspond to the actual situation, the diagnostic standard module is updated.

[0117] For the diagnostic standard update module, if the judgment is inaccurate, the characteristic parameters a of the inaccurate components are saved. i , the distance from the point to the actual centroid is the clustering boundary of the class, and the alarm threshold δ n+1 to update.

[0118] If the faulty component does not correspond to the actual one, the knowledge graph is updated, the relationship between the real graph and the entity in the knowledge graph is updated, the data is retrained, and the knowledge graph is updated.

Claims

1. A wind turbine online automatic diagnosis system based on knowledge graph, characterized by First, the improved Faster RCNN model is used to extract the fault characteristic parameters of the vibration acceleration in the wind turbine, and then the fault characteristics are aggregated to obtain the fault diagnosis criteria; and the fault diagnosis module is combined with the knowledge graph, and the diagnosis results and feedback are combined to propose an online automatic diagnosis system for wind turbines with autonomous learning and reasoning update capabilities.

2. Collect data on the wind turbine drive chain, blades, nacelle, and tower, and transmit it to the local intelligent terminal of each wind turbine for time-frequency conversion. The formula for time-frequency conversion is as follows: in, k=0,1,...,N-1; The output signal's spectrogram is 640x640 in size and in BGR format. The data from each wind turbine’s local intelligent terminal is then transmitted to the knowledge acquisition layer. The knowledge acquisition layer is used for feature extraction and collection of prior knowledge.

3. Perform feature extraction on the data collected from each unit on the local intelligent terminal of the wind turbine. The process is as follows. S1: The vibration analysis staff calculates the fault characteristic frequency of the unit based on the parameters of each component of the unit and marks it in the spectrum. The frequencies marked include: Blade: first-order dancing frequency, first-order shimmying frequency, second-order dancing frequency, second-order shimmying frequency, blade rotation frequency, first-order torsional vibration frequency, second-order torsional vibration frequency; Tower: first-order resonant frequency of the tower, second-order resonant frequency of the tower, first-order torsional vibration frequency, second-order torsional vibration frequency; Spindle: spindle rotation frequency, spindle bearing fault characteristic frequency; Gearbox: gear fault characteristic frequency of each gear in the gearbox, meshing frequency of each gear in the gearbox and its multiples, gearbox bearing fault characteristic frequency and its multiples, and gearbox support structure; Generator: generator rotation frequency and its multiples, generator bearing fault characteristic frequency, generator pole passing frequency, power frequency, etc. S2: Build a feature extraction network model and input the spectrum graph into the network. The process is as follows: S2.1: First, input the image into the feature extraction network to obtain the corresponding feature map. The feature extraction module is shown in Figure 5. "m×n conv, X" is a convolutional layer with a kernel width of m and a kernel height of n, respectively. The number of channels of the convolution is X. For example, "3×3conv, 64" indicates a convolution calculation with a kernel width of 3 and a kernel height of 3 and a kernel number of 64. Pool / 2 represents the maximum pooling layer with a pooling kernel of 2*2. S2.2: The extracted feature map is passed through the RPN (Region Proposal Network) network to generate candidate boxes. The candidate boxes generated by the RPN network are projected onto the feature map to obtain the corresponding feature matrix. The PRN network structure is shown in Figure 4 below: For each 3×3 sliding window of the feature map, we first calculate the position of the center point on the original image, and then generate a series of anchor boxes with this position as the center. K represents the number of anchor boxes, and 2k is the two probabilities generated for each anchor point, one is the probability of being the background, and the other is the probability of being the target feature frequency. If the object in an anchor box is the feature frequency to be detected, then it is the target feature frequency, otherwise it is the background. For each anchor box, four bounding box regression parameters are generated. The regression parameters are also a predicted offset for the center coordinates of the anchor box and an adjustment for the width and height. Through 3×3 convolution, setting the stride to 1, a feature matrix with the same height, width and depth as the original feature map is generated. S2.3: These generated feature matrices are scaled to a uniform size through the RoIPooling layer, and finally scaled to a 7×7 size. The feature matrices are then flattened and passed through the fully connected layer to obtain the predicted probability and bounding box regression parameters. Two 1×1 fully connected layers are connected in parallel on the obtained feature matrix to predict the bounding box regression parameters and categories. S2.4: The loss function of the RPN part consists of two parts, one is the classification loss and the other is the bounding box regression loss. in, where p i Represents the probability that the i-th anchor box is predicted to be the true label; Indicates that the current calculated sample is 1 when it is a positive sample and 0 when it is a negative sample; t i represents the bounding box regression parameters for predicting the i-th anchor box; Represents the regression parameters of the bounding box in the real box corresponding to the i-th anchor box; N cls Indicates the number of samples, N in the present invention cls =256; N reg Indicates the number of anchor box positions. In this invention, N reg =2400; L cls is the classification loss, calculated as follows: L reg is the bounding box regression loss, calculated as follows: Calculate the loss function of the regression boundary to minimize it and obtain the final target image I. After model calculation, the eigenvalue of the target image is: r←Faster RCNN(I) S2.5: After feature extraction, the eigenvalues ​​are estimated. The estimation process is as follows: F1: Corrected eigenvalue r * :r * =r+w In the formula w is the current noise and obeys Gaussian white noise; F2: Define the corrected observation equation C: C * =r * +v * In the formula v * is the current noise, and obeys Gaussian white noise; F3: Correct the correlation parameter P of the current eigenvalue * :P * =P+Q In the formula P is the updated value of the current feature value; Q is the covariance of noise w; F4: Update eigenvalue gain M * :M * =P * (P * +J) -1 In the formula J is the noise v * covariance of (P * +J) -1 For (P * +J); F5: Estimated eigenvalues : In the formula is the correction value of the target image. S2.6: Calculate the loss function of Faster RCNN: L(p,u,t u ,v)=L cls (p,u)+λ[u≥1]L loc (t u ,v) Where P represents the softmax probability distribution predicted by the classifier P = (P0, ..., P K ) P0 represents the probability that the candidate region is the background u represents the corresponding target true category label t u The regression parameters of the corresponding category u predicted by the corresponding bounding box regressor v corresponds to the bounding box regression parameters of the true target (v x ,v y ,v w ,v h ) L cls (p,u)=-logp u , is the classification loss Among them, p u Indicates the probability that the classifier predicts that the current candidate region is category u S3: Input the real-time data of the wind turbine into the trained model to identify the target frequency and frequency size in the model. The target characteristic frequency is the sensitivity factor of the data measurement point under the same working conditions, which is input into the operation and maintenance Internet of Things.

4. Through the operation and maintenance Internet of Things, the sensitive factors of each measuring point under the same working conditions are selected and input into the diagnostic standard generation module to first analyze the characteristic values. The algorithm steps are as follows: T1: Determine the number of clusters to be 2; T2: Randomly select 2 data points from the data set as the initial centroid, denoted as T3: For each sample point a i , assign it to the nearest centroid; T4: For each class center, recalculate the center of the class: in a i is the i-th sample; c i for a i The cluster to which it belongs; is the center of the cluster, i.e. the centroid; M is the total number of samples. T5: Repeat steps T3 and T4. When the centroid no longer changes, stop and output the clustering results. The class with a large number of cases in the judgment class determines that the standard state of the component is normal, otherwise it is abnormal, where the threshold of the normal state is δ.

5. Construct a fault tree of the monitored components based on the condition monitoring knowledge base and existing data, taking the generator as an example, as shown in Figure 3, and calculate the probability of each event occurring. According to the fault tree, a knowledge graph in the form of triples of G = (E, R, S) is constructed, where E = {e1, e2, e3, ...e n } is the collection of views in the knowledge graph, R = {r1, r2, r3, ..., r n } is a collection of relations in the knowledge graph. S = E × R × E represents a triple consisting of entities and relations. The basic form can be entity 1-relationship-entity 2. As shown in Figure 7, the generator fault includes the rotor fault. The algorithm for knowledge graph reasoning is as follows: E1: Consider the knowledge graph as a Markov process with a stationary distribution of π(h), inputting the state transition matrix Q, the number of state transitions n1, the number of samples n2, and the adjacency matrix x t ; E2: Sample the initial state value h0 from any simple probability distribution; E3: For t = 0 to n1 + n2 - 1, E3.1: From the conditional probability distribution Q(h|h t ) Sampling to obtain sample h * ; E3.2: Sample U ∼ [0, 1] from a uniform distribution; E3.3: If Then accept h t+1 =h * ; In the formula, α(h n ,h * )=π(h * )Q(h * ,h n ); E3.4: Otherwise, the transfer is not accepted. t+1 =h t ; E4: Return sample set That is the corresponding sample set that conforms to the stationary distribution. When an indicator in vibration monitoring exceeds the alarm threshold and triggers an alarm, on-site operation and maintenance personnel can check the fault location based on the alarm system and investigate the faulty components and the extent of the fault. After investigation, the operation and maintenance personnel will enter the investigation results into the system. If the faulty component does not correspond, the state monitoring experience database is updated, the knowledge graph structure is updated, and the reasoning algorithm of the above knowledge graph is recalculated to calculate the fault transfer situation. If the fault degree does not correspond to the actual situation, the diagnostic standard module is updated. For the diagnostic standard update module, if the judgment is inaccurate, the characteristic parameters a of the inaccurate components are saved. i , the distance from the point to the actual centroid is the clustering boundary of the class, and the alarm threshold δ n+1 to update. If the faulty component does not correspond to the actual one, the knowledge graph is updated, the relationship between the real graph and the entity in the knowledge graph is updated, the data is retrained, and the knowledge graph is updated.

Images