A method for predicting cracking faults of fan blades based on parallel system
By combining the parallel system of the Inception-Resnet V2 model and Self Attention mechanism, the automation and accuracy of fan blade cracking fault prediction under complex working conditions is solved, real-time judgment and independent learning of fan blade status are realized, and fault prediction is improved accuracy and safety of wind turbines.
Patent Information
- Application Number
- CN202210538049.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-18
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-05-18
AI Technical Summary
The existing methods of fan blade cracking fault prediction lack automation in complex working conditions, model prediction is not accurate enough, and data cannot be collected in real time for judgment, which affects the safety and efficiency of wind turbines.
Using a parallel system-based method, the Inception-Resnet V2 model is combined with the Self Attention mechanism, and through data cleaning and multimodal data processing, a 299×299×3 image is generated for feature extraction and weighted fusion, using the root mean square transfer gradient descent The algorithm optimizes the network model to achieve fast and accurate prediction of fan blade cracking faults.
Real-time judgment and independent learning of the fan blade status are realized, the accuracy and reliability of fault prediction are improved, the blade damage can be prevented earlier and the operation and maintenance costs can be reduced.
Smart Images

Figure FDA0004988338990000031 
Figure FDA0004988338990000032 
Figure FDA0004988338990000033
Abstract
Description
Technical Field
[0001] The invention belongs to the field of prediction of cracking faults of fan blades, and in particular relates to an artificial intelligence prediction method for cracking faults of fan blades based on a parallel system. Background Art
[0002] Wind turbine blades are inevitably damaged when they are operating in harsh natural environments for a long time. As the damage continues to deepen, it will seriously threaten the safe operation of wind turbines. Blade cracking will harm the shaft system, tower and other components of the generator set, and even cause tower collapse. How to detect the status of blades, especially how to quickly and accurately determine the damage status of blades when cracks appear under complex working conditions, is crucial to ensuring the overall efficiency and safety and reliability of wind turbines.
[0003] In addition, in the traditional research on the detection of wind turbine blade damage, the detection methods adopted include ultrasonic detection, infrared imaging detection, vibration abnormality monitoring and acoustic emission detection technologies, but such detection methods are unable to detect wind turbine blades in working state, and have high requirements for the detection environment and equipment. In the current research on the prediction of blade cracking failure of wind turbines in working state, the data collected by the data acquisition and monitoring control system mainly uses the random forest algorithm for feature selection, and then uses the LightGBM or XGBoost machine learning algorithm to train the model, and uses five-fold cross validation to optimize the model, but such methods lack automation means and can only perform mechanical preprocessing on the data. The selection of the five-fold cross validation threshold requires manual debugging, and the model is not accurate enough in predicting the cracking failure of wind turbine blades.
[0004] Therefore, a fan blade fault prediction method with high accuracy that can judge and autonomously learn the data collected in real time on the fan blade status is proposed. Summary of the invention
[0005] The present invention proposes a fan blade cracking fault prediction method based on a parallel system, which combines the Inception-Resnet V2 model, the Self Attention mechanism and the parallel system for fan blade cracking fault prediction. It can more quickly and accurately predict whether a fan blade will crack within a week and prevent the fan blade from being seriously damaged due to continued operation due to hidden dangers. The steps in the use process are:
[0006] Step (1): cleaning the multi-dimensional data of the wind turbine collected by the data acquisition and monitoring control system, including: format standardization, deletion of erroneous data, removal of duplicate data, and removal of unchanged feature dimensions;
[0007] Step (2): After data cleaning, the cleaned fan multi-dimensional data is divided into five types according to different modes: multi-granularity, multi-time scale, multi-source, multi-level, and multi-operating condition. If the fan data dimension after data cleaning and classification is m, the number of real-time multi-dimensional data generated by the fan within 10 minutes is selected as m, and an m×m image is generated, which is then scaled to a 299×299 image. Finally, the 299×299 image is superimposed on itself to obtain a 299×299×3 image.
[0008] The wind turbine multi-dimensional data is divided into five categories according to different modes: multi-granularity, multi-time scale, multi-source, multi-level and multi-operating conditions. The specific categories include:
[0009] According to the multi-granularity, it is divided into: blade tip granular data, blade shell granular data and blade main beam granular data;
[0010] According to multiple time scales, it is divided into: hourly accuracy data for the next day, 30-minute accuracy data for the next day, and 10-minute accuracy data for the next day;
[0011] According to multiple sources, it is divided into: blade status data, inverter status data, generator status data, pitch motor status data, hub, drive and main bearing status data and other status parameters;
[0012] It is divided into multiple levels: wind-related data, energy-related data and temperature-related data;
[0013] According to multiple operating conditions, it is divided into: shutdown state data, low-speed state data, high-speed state data, optimal tip speed ratio stage data, constant speed data and constant power data;
[0014] Step (3): After the image is generated, the image is input into the preset network model for training. The weight w and bias b of the network model loss function are updated using the root mean square transfer gradient descent algorithm. When the error or the number of iterations reaches the threshold, the training stops. The test sample is then substituted into the trained network. The five modal results are then weighted and fused to obtain the prediction result.
[0015] The preset network model includes an Inception-Resnet V2 model and a Self Attention mechanism, wherein the Inception-Resnet V2 is used to extract the generated 299×299×3 image features; and the Self Attention mechanism is used to calculate the weights in the image features to obtain updated image features;
[0016] The specific content of the preset network model is:
[0017] The input is a 299×299×3 image;
[0018] Connect to 1 Stem module;
[0019] Then add 5 Inception-Resnet-A modules;
[0020] It is further divided into 2 branches:
[0021] Branch 1 is a Self Attention mechanism;
[0022] Branch 2 is directly output without processing;
[0023] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0024] Then connect a Reduction-A module;
[0025] Then connect 10 Inception-Resnet-B;
[0026] It is further divided into 2 branches:
[0027] Branch 1 is a Self Attention mechanism;
[0028] Branch 2 is directly output without processing;
[0029] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0030] Then connect a Reduction-B module;
[0031] Then add another Inception-Resnet-C module;
[0032] It is further divided into 2 branches:
[0033] Branch 1 is a Self Attention mechanism;
[0034] Branch 2 is directly output without processing;
[0035] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0036] Then connect to an Average Pooling layer, a Dropout layer, and a Softmax layer in sequence;
[0037] Finally, connect to a classification layer and output the prediction results;
[0038] The specific contents of the Stem module structure are:
[0039] The input is a 299×299×3 image;
[0040] After three 3×3 convolutions;
[0041] It is further divided into 2 branches:
[0042] Branch 1 is a 3×3 convolution process;
[0043] Branch 2 is a 3×3 maximum pooling process;
[0044] Merge branch 1 and branch 2;
[0045] It is further divided into 2 branches:
[0046] Branch 1 undergoes a 1×1 32-channel convolution and a 3×3 32-channel convolution in sequence;
[0047] Branch 2 undergoes a 1×1 convolution with 64 channels, a 7×1 convolution with 64 channels, a 1×7 convolution with 64 channels, and a 3×3 convolution with 96 channels.
[0048] Merge branch 1 with branch 2;
[0049] It is further divided into 2 branches:
[0050] Branch 1 is a 3×3 convolution process;
[0051] Branch 2 is a maximum pooling process;
[0052] Finally, the two branches are merged and output;
[0053] The specific contents of the Inception-Resnet-A module structure are:
[0054] After input, it is divided into 4 branches:
[0055] Branch 1 is the direct output without processing;
[0056] Branch 2 is a 1×1 32-channel convolution process;
[0057] Branch 3 undergoes a 1×1 32-channel convolution process and a 3×3 32-channel convolution process in sequence;
[0058] Branch 4 undergoes a 1×1 32-channel convolution, a 3×3 48-channel convolution, and a 3×3 64-channel convolution in sequence;
[0059] The results of branches 2, 3, and 4 are stacked and convolved again, and then added to the result of branch 1.
[0060] Finally output;
[0061] The specific contents of the Reduction-A module structure are:
[0062] After input, it is divided into 3 branches:
[0063] Branch 1 is a 3×3 maximum pooling process;
[0064] Branch 2 is a 3×3 convolution process;
[0065] Branch 3 undergoes a 1×1 convolution of 256 channels, a 3×3 convolution of 256 channels, and a 3×3 convolution in sequence;
[0066] Then add the results of branches 1, 2, and 3;
[0067] Finally output;
[0068] The specific contents of the Inception-Resnet-B module structure are:
[0069] After input, it is divided into 3 branches:
[0070] Branch 1 is the direct output without processing;
[0071] Branch 2 is a 1×1 192-channel convolution process;
[0072] Branch 3 undergoes a 1×1 convolution with 128 channels, a 1×7 convolution with 160 channels, and a 7×1 convolution with 192 channels.
[0073] The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1;
[0074] Finally output;
[0075] The specific contents of the Reduction-B module structure are:
[0076] After input, it is divided into 4 branches:
[0077] Branch 1 is a 3×3 maximum pooling process;
[0078] Branch 2 undergoes a 1×1 256-channel convolution and a 3×3 convolution in sequence;
[0079] Branch 3 undergoes a 1×1 256-channel convolution process and a 3×3 convolution process in sequence;
[0080] Branch 4 is processed by a 1×1 convolution of 256 channels, a 1×7 convolution of 256 channels, and a 7×1 convolution.
[0081] Then add the results of branches 1, 2, 3, and 4;
[0082] Finally output;
[0083] The specific contents of the Inception-Resnet-C module structure are:
[0084] After input, it is divided into 3 branches:
[0085] Branch 1 is the direct output without processing;
[0086] Branch 2 is a 1×1 192-channel convolution process;
[0087] Branch 3 undergoes a 1×1 convolution of 192 channels, a 1×3 convolution of 224 channels, and a 3×1 convolution of 256 channels in sequence;
[0088] The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1;
[0089] Finally output;
[0090] The specific content of the Self Attention mechanism structure is:
[0091] For the input vector X, multiply it by three matrices W obtained through learning Q , W K and W V The calculation formulas of the three variables Q, K, and V are:
[0092] Q=W Q X(1)
[0093] K=W K X(2)
[0094] V=W V X(3) Where Q, K and V represent Query, Key and Value respectively; Query is equivalent to the information to be queried, Key is equivalent to the database to be queried, and Value is the value obtained by the query;
[0095] The calculation formula of the weight matrix α obtained using softmax is:
[0096]
[0097] Among them, d K is the dimension of Key, K T is the transposed matrix of matrix K;
[0098] The weight matrix α and V are used for weighting, and the calculation formula of the updated image feature weight Attention (Q, K, V) is:
[0099]
[0100] The loss function is the root mean square error loss function, and the calculation formula of the root mean square error loss function L is:
[0101]
[0102] Among them, k is the number of samples, y i and i ′ represents the true value and predicted value of the i-th sample respectively; the smaller the value of the loss function, the closer the true value and predicted value of the network are;
[0103] The RMS transfer gradient descent algorithm accumulates the weighted average of the squared gradients of the weights during the first t iterations. The weight w updated in the tth iteration t , the weighted average of the squared gradient of the bias accumulated in the first t iterations and the bias b updated at the tth iteration t The calculation formulas are:
[0104]
[0105] Where t is the number of iterations, is the weighted average of the squared gradients of the weights accumulated in the first t-1 iterations, is the weighted average of the squared gradient of the bias accumulated in the first t-1 iterations, α is the learning rate, is the gradient of the weight at the t-1th iteration, is the gradient of the bias at the t-1th iteration, ε is a constant added to maintain numerical stability, and w t-1 is the weight updated in the t-1th iteration, b t-1 is the bias of the t-1th iteration update, β is the attenuation factor, which represents the influence of the past gradient on the current gradient;
[0106] Step (4): after obtaining the prediction results, four artificial systems for predicting fan blade cracking faults are established, which have the same structure as the actual system for predicting fan blade cracking faults but different network parameters; these four artificial systems for predicting fan blade cracking faults and the actual system for predicting fan blade cracking faults together constitute a parallel fan blade cracking fault prediction system; through computational experiments on the artificial system for predicting fan blade cracking faults, the development and evolution of the actual system scenario for predicting fan blade cracking faults is guided, and the changes in the actual system scenario for predicting fan blade cracking faults are fed back to the artificial system for predicting fan blade cracking faults in real time, and the difference between the artificial system for predicting fan blade cracking faults and the actual system for predicting fan blade cracking faults is selected to compare the operation of the artificial system for predicting fan blade cracking faults that obtains the best prediction result, and then the parameters of the actual system for predicting fan blade cracking faults are adjusted and optimized to obtain a more optimized fault prediction result;
[0107] The fan blade crack fault prediction actual system includes: performing data cleaning on the fan multi-dimensional data collected by using a data acquisition and monitoring control system; after the data cleaning, the fan multi-dimensional data after the data cleaning is divided into 5 categories according to different modes to generate images; after the image is generated, the image is input into a preset network model to obtain a prediction result.
[0108] Compared with the prior art, the present invention has the following advantages and effects:
[0109] (1) The present invention processes 299×299×3 images of different modes through a preset network model. The Inception-Resnet V2 model includes one or more 1×1, 3×3, 1×3, 3×1, 1×7 and 7×1 convolution kernel branches. The multi-core structure can extract and learn features of 299×299×3 images of different modes. At the same time, this multi-scale convolution makes it unnecessary to manually design the convolution kernel size or determine whether to create convolution layers and pooling layers in the module, thereby reducing the impact of human factors on the network. At the same time, the Resnet module introduced in the Inception-Resnet V2 model makes the gradient larger, making the model converge faster. The Self Attention mechanism is used to calculate the weights in the image features of different layers. The feature maps of different layers are weighted and fused through the Self Attention mechanism to obtain a multi-scale feature map, thereby obtaining the target area that needs to be focused on and obtaining the focus of attention. More detailed information of the target that needs to be focused on can be obtained, while suppressing other useless information, and more effectively extracting features, making the prediction results of fan blade cracking faults more accurate.
[0110] (2) The present invention establishes a parallel fan blade cracking fault prediction system to address complexity issues, making complex systems computable, experimental, and verifiable, and making complex controls quantifiable, analytical, and visual. The parallel fan blade cracking fault prediction system uses computational experiments to fully explore different scenarios in the fan blade cracking fault prediction artificial system. Through advanced artificial intelligence and big data methods, the system is explored, learned, and predicted, accumulating knowledge and experience in complex system management and control problems, solving unsolvable problems encountered in the real world, and truly addressing the challenges of inability to accurately predict, difficult to split and restore, and inability to repeat experiments in actual complex systems. The management and control scheme of the fan blade cracking fault prediction artificial system can be used as a backup plan for the actual fan blade cracking fault prediction system. Compared with other fan blade cracking fault prediction systems, the parallel fan blade cracking fault prediction system is more reliable and adaptable. Through computational experiments on the artificial system for predicting fan blade cracking faults, the development and evolution of the actual system scenarios for predicting fan blade cracking faults are guided, and the changes in the actual system scenarios are fed back to the artificial system for predicting fan blade cracking faults in real time. The differences in the operation of the artificial system for predicting fan blade cracking faults that obtains the best prediction results and the actual system for predicting fan blade cracking faults are selected to adjust and optimize the parameters of the actual system for predicting fan blade cracking faults, thereby realizing the integration of virtual and real interaction and rapid optimization, and being able to obtain more accurate fault prediction results.
[0111] (3) By analyzing the multi-dimensional data of the wind turbine in operation, the present invention can more quickly and accurately predict whether the wind turbine blades will crack within a week, which is beneficial for operation and maintenance personnel to maintain the blades at an earlier time and prevent the blades from being seriously damaged due to continued operation due to hidden dangers, thereby more effectively reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0112] Figure 1 It is a flow chart of a fan blade cracking fault prediction method based on a parallel system of the present invention.
[0113] Figure 2 It is a classification diagram of the multi-dimensional data of the fan according to different modes according to the method of the present invention.
[0114] Figure 3 It is the overall structure diagram of the network model preset in the present invention.
[0115] Figure 4 It is a structural diagram of the Stem module in the present invention.
[0116] Figure 5 This is the structure diagram of the Inception-Resnet-A module in the present invention.
[0117] Figure 6It is the structural diagram of the Reduction-A module in the present invention.
[0118] Figure 7 This is the structure diagram of the Inception-Resnet-B module in the present invention.
[0119] Figure 8 It is the structural diagram of the Reduction-B module in the present invention.
[0120] Fig. 9 This is the structure diagram of the Inception-Resnet-C module in the present invention.
[0121] Fig.10 It is a structural diagram of the Self Attention mechanism in the present invention.
[0122] Fig.11 It is a basic framework diagram of the parallel control of the fan blade cracking fault prediction system in the present invention.
[0123] Fig.12 It is the overall framework of the parallel fan blade cracking fault prediction system in the present invention.
[0124] Fig.13 It is a correlation diagram of the parallel fan blade cracking fault prediction system and control and management in the present invention. DETAILED DESCRIPTION
[0125] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0126] The fan blade cracking fault prediction method based on the parallel system proposed by the present invention is described in detail as follows in conjunction with the accompanying drawings:
[0127] Figure 1 It is a flow chart of a fan blade cracking fault prediction method based on a parallel system of the present invention.
[0128] It is important to emphasize that: Figure 1 The flowchart shown is only a preferred embodiment. Those skilled in the art should know that any embodiment constructed around the concept of the present invention should not deviate from the scope covered by the following technical solutions:
[0129] The multi-dimensional data of the fan collected by the data acquisition and monitoring control system is cleaned; after data cleaning, the cleaned multi-dimensional data of the fan is divided into 5 categories according to different modes to generate images; after the image is generated, the image is input into the preset network model to obtain the prediction result; after obtaining the prediction result, a parallel fan blade cracking fault prediction system is constructed, and the difference in operation between the artificial system for predicting fan blade cracking faults and the actual system for predicting fan blade cracking faults is selected to obtain the best prediction result, and then the parameters of the actual system for predicting fan blade cracking faults are adjusted and optimized to obtain a more optimized fault prediction result.
[0130] Figure 2 This is a classification diagram of wind turbine multi-dimensional data according to different modes of the method of the present invention. The wind turbine multi-dimensional data of the same group is extracted and divided into five categories according to different modes: multi-granularity, multi-time scale, multi-source, multi-level and multi-operating conditions. The specific classification of these five categories includes:
[0131] According to the multi-granularity, it is divided into: blade tip granular data, blade shell granular data and blade main beam granular data;
[0132] According to multiple time scales, it is divided into: hourly accuracy data for the next day, 30-minute accuracy data for the next day, and 10-minute accuracy data for the next day;
[0133] According to multiple sources, it is divided into: blade status data, inverter status data, generator status data, pitch motor status data, hub, drive and main bearing status data and other status parameters;
[0134] It is divided into multiple levels: wind-related data, energy-related data and temperature-related data;
[0135] It is divided into multiple operating conditions: shutdown state data, low-speed state data, high-speed state data, optimal tip speed ratio stage data, constant speed data and constant power data.
[0136] Figure 3 It is the overall structure diagram of the preset network model in the present invention. The specific contents of the preset network model are:
[0137] The input is a 299×299×3 image;
[0138] Connect to 1 Stem module;
[0139] Then add 5 Inception-Resnet-A modules;
[0140] It is further divided into 2 branches:
[0141] Branch 1 is a Self Attention mechanism;
[0142] Branch 2 is directly output without processing;
[0143] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0144] Then connect a Reduction-A module;
[0145] Then connect 10 Inception-Resnet-B;
[0146] It is further divided into 2 branches:
[0147] Branch 1 is a Self Attention mechanism;
[0148] Branch 2 is directly output without processing;
[0149] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0150] Then connect a Reduction-B module;
[0151] Then add another Inception-Resnet-C module;
[0152] It is further divided into 2 branches:
[0153] Branch 1 is a Self Attention mechanism;
[0154] Branch 2 is directly output without processing;
[0155] Then branch 1 is multiplied by branch 2 and the output is obtained;
[0156] Then connect to an Average Pooling layer, a Dropout layer, and a Softmax layer in sequence;
[0157] Finally, a classification layer is connected to output the prediction results.
[0158] Figure 4 It is the structure diagram of the Stem module in the present invention. The specific contents of the Stem module structure are:
[0159] The input is a 299×299×3 image;
[0160] After three 3×3 convolutions;
[0161] It is further divided into 2 branches:
[0162] Branch 1 is a 3×3 convolution process;
[0163] Branch 2 is a 3×3 maximum pooling process;
[0164] Merge branch 1 and branch 2;
[0165] It is further divided into 2 branches:
[0166] Branch 1 undergoes a 1×1 32-channel convolution and a 3×3 32-channel convolution in sequence;
[0167] Branch 2 undergoes a 1×1 convolution with 64 channels, a 7×1 convolution with 64 channels, a 1×7 convolution with 64 channels, and a 3×3 convolution with 96 channels.
[0168] Merge branch 1 with branch 2;
[0169] It is further divided into 2 branches:
[0170] Branch 1 is a 3×3 convolution process;
[0171] Branch 2 is a maximum pooling process;
[0172] Finally, the two branches are merged and output;
[0173] Figure 5 It is the Inception-Resnet-A module structure diagram of the present invention. The specific contents of the Inception-Resnet-A module structure are:
[0174] After input, it is divided into 4 branches:
[0175] Branch 1 is the direct output without processing;
[0176] Branch 2 is a 1×1 32-channel convolution process;
[0177] Branch 3 undergoes a 1×1 32-channel convolution process and a 3×3 32-channel convolution process in sequence;
[0178] Branch 4 undergoes a 1×1 32-channel convolution, a 3×3 48-channel convolution, and a 3×3 64-channel convolution in sequence;
[0179] The results of branches 2, 3, and 4 are stacked and convolved again, and then added to the result of branch 1.
[0180] Final output.
[0181] Figure 6 : is the Reduction-A module structure diagram of the present invention. The specific contents of the Reduction-A module structure are:
[0182] After input, it is divided into 3 branches:
[0183] Branch 1 is a 3×3 maximum pooling process;
[0184] Branch 2 is a 3×3 convolution process;
[0185] Branch 3 undergoes a 1×1 convolution of 256 channels, a 3×3 convolution of 256 channels, and a 3×3 convolution in sequence;
[0186] Then add the results of branches 1, 2, and 3;
[0187] Final output.
[0188] Figure 7 It is the structure diagram of the Inception-Resnet-B module in the present invention. The specific contents of the Inception-Resnet-B module structure are:
[0189] After input, it is divided into 3 branches:
[0190] Branch 1 is the direct output without processing;
[0191] Branch 2 is a 1×1 192-channel convolution process;
[0192] Branch 3 undergoes a 1×1 convolution with 128 channels, a 1×7 convolution with 160 channels, and a 7×1 convolution with 192 channels.
[0193] The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1;
[0194] Final output.
[0195] Figure 8 : is the Reduction-B module structure diagram of the present invention. The specific contents of the Reduction-B module structure are:
[0196] After input, it is divided into 4 branches:
[0197] Branch 1 is a 3×3 maximum pooling process;
[0198] Branch 2 undergoes a 1×1 256-channel convolution and a 3×3 convolution in sequence;
[0199] Branch 3 undergoes a 1×1 256-channel convolution process and a 3×3 convolution process in sequence;
[0200] Branch 4 is processed by a 1×1 convolution of 256 channels, a 1×7 convolution of 256 channels, and a 7×1 convolution.
[0201] Then add the results of branches 1, 2, 3, and 4;
[0202] Finally output;
[0203] Fig. 9 It is the Inception-Resnet-C module structure diagram of the present invention. The specific content of the Inception-Resnet-C module structure is:
[0204] After input, it is divided into 3 branches:
[0205] Branch 1 is the direct output without processing;
[0206] Branch 2 is a 1×1 192-channel convolution process;
[0207] Branch 3 undergoes a 1×1 convolution of 192 channels, a 1×3 convolution of 224 channels, and a 3×1 convolution of 256 channels in sequence;
[0208] The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1;
[0209] Final output.
[0210] Fig.10 It is a diagram of the structure of the Self Attention mechanism in the present invention. The specific contents of the Self Attention mechanism structure are:
[0211] For the input vector X, multiply it by three matrices W obtained through learning Q , W K and W V The calculation formulas of the three variables Q, K, and V are:
[0212] Q=W Q X
[0213] K=W K X
[0214] V=W V X
[0215] Among them, Q, K and V represent Query, Key and Value respectively; Query is equivalent to the information to be queried, Key is equivalent to the database to be queried, and Value is the value obtained by the query;
[0216] The calculation formula of the weight matrix α obtained using softmax is:
[0217]
[0218] Among them, d K is the dimension of Key, K Tis the transposed matrix of matrix K;
[0219] The weight matrix α and V are used for weighting, and the calculation formula of the updated image feature weight Attention (Q, K, V) is:
[0220]
[0221] Fig.11 This is the basic framework diagram of the parallel control of the fan blade cracking fault prediction system in the present invention. The parallel fan blade cracking fault prediction system includes a fan blade cracking fault prediction actual system and a fan blade cracking fault prediction artificial system, which operates in three basic modes: learning and training, experiment and evaluation, and management and control. The specific description is as follows:
[0222] (I) Experiment and evaluation: In this process, the actual system for predicting the cracking of fan blades operates on the basis of rules, while the artificial system for predicting the cracking of fan blades converts data into knowledge, which can be understood as a knowledge-driven system. The artificial system for predicting the cracking of fan blades is mainly used to conduct computational experiments, analyze and understand the experimental behaviors and reactions of four different artificial systems for predicting the cracking of fan blades, and evaluate the effects of different prediction results as a basis for selecting and supporting management and control decisions.
[0223] (ii) Learning and training: In this process, the fan blade cracking fault prediction artificial system is mainly used as a center for learning and training management and control of the entire parallel fan blade cracking fault prediction system. By appropriately connecting and combining the actual fan blade cracking fault prediction system with the fan blade cracking fault prediction artificial system, the relevant personnel who manage and control the actual complex system can quickly grasp the various conditions of the system and the corresponding actions. When conditions permit, the artificial system should be operated with a management and control system equivalent to the actual one in order to obtain better real results. At the same time, the management and control scheme of the fan blade cracking fault prediction artificial system can also be used as a backup plan for the actual fan blade cracking fault prediction system to increase its operational reliability and resilience.
[0224] (III) Management and control: In this process, the fan blade cracking fault prediction artificial system attempts to simulate the actual complex system as much as possible and predict its behavior, so as to provide a basis for finding effective solutions for the actual system or improving the current solution. Furthermore, by observing the difference between the actual state of the fan blade cracking fault prediction system and the state of the fan blade cracking fault prediction artificial system, an error feedback signal is generated, and the evaluation method or parameters of the fan blade cracking fault prediction artificial system are corrected to reduce the difference, and then a new round of optimization and evaluation is analyzed.
[0225] Parallel control targets complex problems, especially complex systems involving people and society, and restores the model from a system analyzer to a data generator, making complex systems computable, experimental, and verifiable, and making complex control quantifiable, analytical, and visual. Through the parallel control of the actual system for predicting fan blade cracking faults and the artificial system for predicting fan blade cracking faults, the performance of the system can be continuously optimized, and ultimately a more accurate prediction result for fan blade cracking faults can be output.
[0226] Fig.12 It is the overall framework of the parallel fan blade cracking fault prediction system in the present invention. First, the fan multi-dimensional data collected by the data acquisition and monitoring control system is cleaned. Then the cleaned data is classified according to different modes (X1~X5), the number of real-time multi-dimensional data generated by the fan within 10 minutes is selected as m, and an m×m image is generated, which is then scaled to a 299×299 image, and the 299×299 image is superimposed on itself to obtain a 299×299×3 image (Y1~Y5); after the image is generated, each modal image (Y1~Y5) is input into the preset network model for training and tuning, and then the five modal results are weighted and fused to obtain the prediction result. Finally, four fan blade cracking fault prediction artificial systems with the same structure as the actual system for fan blade cracking fault prediction but different network parameters are established; these four fan blade cracking fault prediction artificial systems and the actual system for fan blade cracking fault prediction together constitute a parallel fan blade cracking fault prediction system.
[0227] Fig.13 This is the correlation diagram between the parallel fan blade cracking fault prediction system and control and management in the present invention. Using the knowledge automation method, the big data of the artificial system virtual space is deeply developed and intellectually mined to effectively deal with uncertainty, diversity and complexity. From the perspective of knowledge automation, the workflow of the parallel fan blade cracking fault prediction system can correspond to the various stages of control and management as follows:
[0228] 1) Knowledge identification: When receiving multi-dimensional data, the parallel wind turbine blade cracking fault prediction system first analyzes the wind turbine status, and then retrieves the encapsulated knowledge through the knowledge identification status to confirm whether relevant knowledge exists within the system.
[0229] 2) Knowledge creation: If no relevant knowledge corresponding to the prediction result can be found through knowledge retrieval in the identification stage, or the existing knowledge is not comprehensive enough, it may trigger the need to create or acquire new knowledge. The parallel wind turbine blade cracking fault prediction system analyzes the data source and extracts knowledge to obtain new knowledge, and confirms whether the new knowledge is valuable through evaluation.
[0230] 3) Knowledge storage: The new knowledge obtained in the knowledge creation phase will be stored in the knowledge base and used for knowledge retrieval in the knowledge identification phase. 4) Knowledge sharing: In the parallel wind turbine blade cracking fault prediction system, knowledge sharing refers to sharing the internal knowledge of the wind turbine blade cracking fault prediction artificial system with the wind turbine blade cracking fault prediction actual system. This process is completed by generating prediction results.
[0231] 5) Knowledge usage: Once the prediction results are obtained, the actual wind blade cracking fault prediction system can use them to solve problems and make decisions.
[0232] 6) Knowledge learning: Through the feedback of the actual system using knowledge, and based on the fan blade cracking fault prediction evaluation method, it can serve as the basis for creating new knowledge or improving existing knowledge in the parallel fan blade cracking fault prediction system. If the knowledge is evaluated as valuable, it will enter the improvement stage; if the knowledge is evaluated as insufficient or incomplete, it needs to return to the knowledge identification or creation stage.
[0233] 7) Knowledge Refinement: In the refinement phase, new value of knowledge is identified or created and is added or updated.
[0234] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A fan blade cracking fault prediction method based on a parallel system, characterized in that: The Inception-Resnet V2 model, the Self Attention mechanism, and the parallel system are combined to predict cracking failures in wind turbine blades. This can more quickly and accurately predict whether a wind turbine blade will crack within a week and prevent serious damage to the wind turbine blades caused by continued operation due to hidden dangers. The steps in use are: Step (1): cleaning the multi-dimensional data of the wind turbine collected by the data acquisition and monitoring control system, including: format standardization, deletion of erroneous data, removal of duplicate data, and removal of unchanged feature dimensions; Step (2): After data cleaning, the cleaned fan multi-dimensional data is divided into five types according to different modes: multi-granularity, multi-time scale, multi-source, multi-level, and multi-operating condition. If the fan data dimension after data cleaning and classification is m, the number of real-time multi-dimensional data generated by the fan within 10 minutes is selected as m, and an m×m image is generated, which is then scaled to a 299×299 image. Finally, the 299×299 image is superimposed on itself to obtain a 299×299×3 image. The wind turbine multi-dimensional data is divided into five categories according to different modes: multi-granularity, multi-time scale, multi-source, multi-level and multi-operating conditions. The specific categories include: According to the multi-granularity, it is divided into: blade tip granular data, blade shell granular data and blade main beam granular data; According to multiple time scales, it is divided into: hourly accuracy data for the next day, 30-minute accuracy data for the next day, and 10-minute accuracy data for the next day; According to multiple sources, it is divided into: blade status data, inverter status data, generator status data, pitch motor status data, hub, drive and main bearing status data and other status parameters; It is divided into multiple levels: wind-related data, energy-related data and temperature-related data; According to multiple operating conditions, it is divided into: shutdown state data, low-speed state data, high-speed state data, optimal tip speed ratio stage data, constant speed data and constant power data; Step (3): After the image is generated, the image is input into the preset network model for training. The weight w and bias b of the network model loss function are updated using the root mean square transfer gradient descent algorithm. When the error or the number of iterations reaches the threshold, the training stops. The test sample is then substituted into the trained network. The five modal results are then weighted and fused to obtain the prediction result. The preset network model includes an Inception-Resnet V2 model and a Self Attention mechanism, wherein the Inception-Resnet V2 is used to extract the generated 299×299×3 image features; and the Self Attention mechanism is used to calculate the weights in the image features to obtain updated image features; The specific content of the preset network model is: The input is a 299×299×3 image; Connect to 1 Stem module; Then add 5 Inception-Resnet-A modules; It is further divided into 2 branches: Branch 1 is a Self Attention mechanism; Branch 2 is directly output without processing; Then branch 1 is multiplied by branch 2 and the output is obtained; Then connect a Reduction-A module; Then connect 10 Inception-Resnet-B; It is further divided into 2 branches: Branch 1 is a Self Attention mechanism; Branch 2 is directly output without processing; Then branch 1 is multiplied by branch 2 and the output is obtained; Then connect a Reduction-B module; Then add another Inception-Resnet-C module; It is further divided into 2 branches: Branch 1 is a Self Attention mechanism; Branch 2 is directly output without processing; Then branch 1 is multiplied by branch 2 and the output is obtained; Then connect to an Average Pooling layer, a Dropout layer, and a Softmax layer in sequence; Finally, connect to a classification layer and output the prediction results; The specific contents of the Stem module structure are: The input is a 299×299×3 image; After three 3×3 convolutions; It is further divided into 2 branches: Branch 1 is a 3×3 convolution process; Branch 2 is a 3×3 maximum pooling process; Merge branch 1 and branch 2; It is further divided into 2 branches: Branch 1 undergoes a 1×1 32-channel convolution and a 3×3 32-channel convolution in sequence; Branch 2 undergoes a 1×1 convolution with 64 channels, a 7×1 convolution with 64 channels, a 1×7 convolution with 64 channels, and a 3×3 convolution with 96 channels. Merge branch 1 with branch 2; It is further divided into 2 branches: Branch 1 is a 3×3 convolution process; Branch 2 is a maximum pooling process; Finally, the two branches are merged and output; The specific contents of the Inception-Resnet-A module structure are: After input, it is divided into 4 branches: Branch 1 is directly output without processing; Branch 2 is a 1×1 32-channel convolution process; Branch 3 undergoes a 1×1 32-channel convolution process and a 3×3 32-channel convolution process in sequence; Branch 4 undergoes a 1×1 32-channel convolution, a 3×3 48-channel convolution, and a 3×3 64-channel convolution in sequence; The results of branches 2, 3, and 4 are stacked and convolved again, and then added to the result of branch 1. Finally output; The specific contents of the Reduction-A module structure are: After input, it is divided into 3 branches: Branch 1 is a 3×3 maximum pooling process; Branch 2 is a 3×3 convolution process; Branch 3 undergoes a 1×1 convolution of 256 channels, a 3×3 convolution of 256 channels, and a 3×3 convolution in sequence; Then add the results of branches 1, 2, and 3; Finally output; The specific contents of the Inception-Resnet-B module structure are: After input, it is divided into 3 branches: Branch 1 is directly output without processing; Branch 2 is a 1×1 192-channel convolution process; Branch 3 undergoes a 1×1 convolution with 128 channels, a 1×7 convolution with 160 channels, and a 7×1 convolution with 192 channels. The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1; Finally output; The specific contents of the Reduction-B module structure are: After input, it is divided into 4 branches: Branch 1 is a 3×3 maximum pooling process; Branch 2 undergoes a 1×1 256-channel convolution and a 3×3 convolution in sequence; Branch 3 undergoes a 1×1 256-channel convolution process and a 3×3 convolution process in sequence; Branch 4 is processed by a 1×1 convolution with 256 channels, a 1×7 convolution with 256 channels, and a 7×1 convolution. Then add the results of branches 1, 2, 3, and 4; Finally output; The specific contents of the Inception-Resnet-C module structure are: After input, it is divided into 3 branches: Branch 1 is the direct output without processing; Branch 2 is a 1×1 192-channel convolution process; Branch 3 undergoes a 1×1 convolution of 192 channels, a 1×3 convolution of 224 channels, and a 3×1 convolution of 256 channels in sequence; The results of branches 2 and 3 are stacked and convolved again, and then added to the result of branch 1; Finally output; The specific content of the Self Attention mechanism structure is: For the input vector X, multiply it by three matrices W obtained through learning Q , W K and W V The calculation formulas of the three variables Q, K, and V are: Q=W Q X(1)K=W K X(2)V=W V X(3) Where Q, K and V represent Query, Key and Value respectively; Query is equivalent to the information to be queried, Key is equivalent to the database to be queried, and Value is the value obtained by the query; The calculation formula of the weight matrix α obtained using softmax is: Among them, d K is the dimension of Key, K T is the transposed matrix of matrix K; The weight matrix α and V are used for weighting, and the calculation formula of the updated image feature weight Attention (Q, K, V) is: The loss function is the root mean square error loss function, and the calculation formula of the root mean square error loss function L is: Among them, k is the number of samples, y i and i ′ represents the true value and predicted value of the i-th sample respectively; the smaller the value of the loss function, the closer the true value and predicted value of the network are; The RMS transfer gradient descent algorithm accumulates the weighted average of the squared gradients of the weights during the first t iterations. The weight w updated in the tth iteration t , the weighted average of the squared gradient of the bias accumulated in the first t iterations and the bias b updated at the tth iteration t The calculation formulas are: Where t is the number of iterations, is the weighted average of the squared gradients of the weights accumulated in the first t-1 iterations, is the weighted average of the squared gradient of the bias accumulated in the first t-1 iterations, α is the learning rate, is the gradient of the weight at the t-1th iteration, is the gradient of the bias at the t-1th iteration, ε is a constant added to maintain numerical stability, and w t-1 is the weight updated in the t-1th iteration, b t-1 is the bias of the t-1th iteration update, β is the attenuation factor, which represents the influence of the past gradient on the current gradient; Step (4): after obtaining the prediction results, four artificial systems for predicting fan blade cracking faults are established, which have the same structure as the actual system for predicting fan blade cracking faults but different network parameters; these four artificial systems for predicting fan blade cracking faults and the actual system for predicting fan blade cracking faults together constitute a parallel fan blade cracking fault prediction system; through computational experiments on the artificial system for predicting fan blade cracking faults, the development and evolution of the actual system scenario for predicting fan blade cracking faults is guided, and the changes in the actual system scenario for predicting fan blade cracking faults are fed back to the artificial system for predicting fan blade cracking faults in real time, and the difference between the artificial system for predicting fan blade cracking faults and the actual system for predicting fan blade cracking faults is selected to compare the operation of the artificial system for predicting fan blade cracking faults that obtains the best prediction result, and then the parameters of the actual system for predicting fan blade cracking faults are adjusted and optimized to obtain a more optimized fault prediction result; The fan blade crack fault prediction actual system includes: performing data cleaning on the fan multi-dimensional data collected by using a data acquisition and monitoring control system; after the data cleaning, the fan multi-dimensional data after the data cleaning is divided into 5 categories according to different modes to generate images; after the image is generated, the image is input into a preset network model to obtain a prediction result.
