Deep learning-based tower structure damage identification method and system

Abstract

The invention relates to a tower structure damage identification method based on deep learning, and the method comprises the following steps: building a data analysis model, carrying out the modal analysis, and extracting the dynamic characteristic parameters of a structure; the acceleration time history of the vertical distribution measuring points under each damage working condition when the structure is excited by the white noise of the substrate is extracted in batches as damage data; creating a data set, and dividing and marking the data set; building an LCNN model for structural damage identification based on a convolutional neural network CNN and a long-short term memory network LSTM; using the training set to train parameters of the LCNN model, using the verification set to verify each training result, and finally using the test set to test the trained model; and carrying out damage identification and positioning on the label-free data by using the trained model. The method has the technical advantages of being applicable to small damage and easy to extract structural dynamic parameters, and can be widely applied to damage identification of various tower structures.

Description

Technical Field

[0001] The present application relates to a method and system for identifying tower structure damage based on deep learning. Background Art

[0002] The structural damage modes of tower structures such as wind power tower barrels mainly include corrosion, cracks, fatigue damage, local yield and deformation, loosening or fracture of connection parts, foundation settlement or inclination, resonance phenomena, thermal expansion and contraction caused by temperature changes, ice and snow accumulation, and dynamic loads generated by the operation of wind turbines. These damages may lead to a decrease in the stability of the tower barrel, affect the normal operation of the wind turbine unit, and may cause the failure of the wind power tower barrel in severe cases. For damages with no obvious phenomena but great harmfulness, general visual and detection means are extremely insensitive to them. The means based on dynamic characteristics have high requirements for data quality and rely on the prior distribution of specific structures, and there are limitations in the damage identification of tower barrel structures. The data processing method is complex and the robustness is poor. Therefore, the means based on dynamic characteristics are not applicable to the damage identification of tower barrel structures.

[0003] To sum up, the existing tower structure damage detection means have limitations and complexity, are less sensitive to minor damages, and it is difficult to extract structural dynamic parameters, resulting in problems such as low damage identification accuracy and complex operation process; the existing structural damage identification methods rely on the prior information of specific structures and have great limitations in the damage identification of other tower structures, so they do not have good applicability. Summary of the Invention

[0004] In order to overcome the defects of the above-mentioned prior art, the present application provides a method and system for identifying tower structure damage based on deep learning, which is applicable to minor damages, easy to extract structural dynamic parameters, does not rely on the prior information of specific structures, and can be widely applicable to the damage identification of various tower structures.

[0005] The present application relates to a method for identifying tower structure damage based on deep learning, and the method includes the following steps:

[0006] Step S1: Establish a data analysis model of the tower structure, perform modal analysis, and extract the dynamic characteristic parameters of the structure;

[0007] Step S2: Batch extract the acceleration time history of the vertical distribution measurement points of the structure when it is subjected to base white noise excitation under each damage condition as damage data;

[0008] Step S3: Create a data set, divide and label the data set, and divide the data set into a training set, a validation set, and a test set;

[0009] Step S4: Build an LCNN model for structural damage identification based on the convolutional neural network CNN and the long short-term memory network LSTM;

[0010] Step S5: Use the training set to train the parameters of the LCNN model, use the validation set to verify each training result, and finally use the test set to test the trained model;

[0011] Step S6: Use the trained model to identify and locate damages in unlabeled data.

[0012] Among them, in step S2, the following steps may be included:

[0013] Step S21: Layer the structure and select the structural load-bearing plate as the target damaged component;

[0014] Step S22: Reduce the stiffness of the target damaged component to different degrees to simulate damages of different degrees;

[0015] Step S23: Extract the acceleration time history of the vertical distribution measuring points when the structure is subjected to base white noise excitation under each damage condition as damage data, and select the structural layer where the target damaged component is located as the corresponding label data;

[0016] Step S24: Use the script tool to call the data analysis model to quickly generate data in batches.

[0017] Among them, in step S3, the following steps may be included:

[0018] Step S31: Split the data generated in step S2 according to the specified sample shape;

[0019] Step S32: Set the damaged layer of the data where each sample is located as the label of the sample.

[0020] Among them, in step S4, the following steps may be included:

[0021] Step S41: Build CNN and LSTM sub-models respectively;

[0022] Step S42: Set parameters, calculate the training loss using the cross-entropy damage function, perform model optimization using the Adam optimization algorithm, set the learning rate, and configure GPU parameters to accelerate training.

[0023] Furthermore, in step S41, the following steps may also be included:

[0024] Step S411: Build a CNN model, use three consecutive one-dimensional convolutional layers to extract features from the data, specify the output dimensions and convolutional kernel heights of the three convolutional layers respectively, where each convolutional layer is followed by a batch normalization layer, an activation layer and a Dropout layer, select the Relu function for the activation layer, and specify the Dropout ratio;

[0025] Step S412: Build an LSTM model, specifying the number of hidden layers, the number of output dimensions, and the Dropout ratio;

[0026] Step S413: Combine the final outputs of the two sub-models and input them into a fully connected layer. The number of input channels of the fully connected layer is the sum of the final output channels of the two sub-models, and the number of output channels of the fully connected layer is the number of final categories;

[0027] Among them, in step S5, the following steps may be included:

[0028] Step S51: Train the model, set the batch size for a single training and the number of training times, set the model to the training state, input the training set into the LCNN model. In each training round, obtain the output classification result through forward propagation, calculate and record the loss between the output value and the true value, update the model parameters through backpropagation, and record and output the training loss and training parameters;

[0029] Step S52: Validate the model, set the model to the evaluation state, input the validation set into the model trained each time after each training, calculate and record the accuracy in each validation. When the accuracy is greater than the highest accuracy in the past training process, save the model parameters at this training time to obtain the LCNN model;

[0030] Step S53: Test the model, load the model with the highest accuracy during the training process, set the model to the evaluation state, input the test set into the trained model, and calculate the test accuracy on the test set as the model performance evaluation index.

[0031] Among them, in step S6, the following steps may be included:

[0032] Step S61: Load the trained model, set the model to the evaluation state, input the pre-prepared unlabeled data into the model, and calculate the model output result corresponding to the data, which is the layer where the damage is located.

[0033] This application also relates to a tower structure damage identification method based on deep learning, including a digital modeling module, a data acquisition module, a data division and dataset creation module, a deep learning model building module, a deep learning model training module, and a damage identification module;

[0034] The digital modeling module establishes a digital model of the tower structure, groups the components and stratifies the structure, and sends the stratification result to the data acquisition module;

[0035] The data acquisition module reduces the stiffness of each layer of components obtained by the digital modeling module to different degrees to simulate damage, extracts the acceleration time history of the vertical distribution measurement points of the tower structure under the excitation of base white noise in each damage condition as damage data, and sends the data to the data partitioning, labeling, and dataset creation module;

[0036] The data partitioning and dataset creation module saves the data collected by the data acquisition module into the dataset, slices the time history data in the time dimension to generate more samples, uses the layer number where the damage of the working condition to which the sample data belongs as its label, randomly selects the sample data and the corresponding label data as the training set, validation set, and test set, and loads the dataset into the deep learning model training module;

[0037] The deep learning model building module builds an LCNN model for damage identification using the dual-channel mode of LSTM and CNN, and sends the built model to the deep learning training module;

[0038] The deep learning training module uses the built LCNN model to train the model on the training set to update the parameters, validates the training results each time on the validation set to adjust the parameters, tests the finally trained model on the test set, and sends the finally trained model to the damage identification module;

[0039] The damage identification module uses the trained LCNN model to identify the test data with unknown labels and outputs the location of the damage.

[0040] A tower structure damage identification method and system based on deep learning according to the present application has the following technical advantages:

[0041] (1) The present application collects the acceleration time history of each measurement point under various working conditions in the numerical model and the actual structure in the vertical distribution, uses the original data as the input data, constructs a complete dataset for the training, validation, and testing of the damage identification algorithm through the collected working conditions, covering various possible damage situations, without data conversion, ensuring no information loss on the one hand and saving the complex procedures of data engineering on the other hand;

[0042] (2) The present application uses a deep learning model to automatically find features from the data and perform damage identification based on these features, updates the model parameters through the training set, corrects the training direction each time through the validation set, tests the final training effect of the model through the test set, and performs unknown damage identification through unlabeled data, accurately, effectively, and quickly learning the damage characteristics of the wind turbine tower structure, so as to achieve accurate identification and positioning of the damage. Description of the Drawings

[0043] Figure 1Schematic flow chart of a tower structure damage identification method based on deep learning proposed in this application;

[0044] Figure 2 Schematic structure diagram of the LCNN model proposed in this application. Detailed implementation manners

[0045] To make the objectives, technical solutions and advantages of this application clearer and more understandable, the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be noted that, without conflict, the embodiments and features in the embodiments of this application can be combined arbitrarily with each other. In this document, the term "including", "comprising" or any other variant thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0046] This application relates to a tower structure damage identification method based on deep learning. The method includes steps such as digital modeling, data generation, data set creation, model construction, model training and damage identification. This application first establishes a numerical model of a wind turbine tower, extracts the acceleration time history of vertical distribution measurement points when the structure is subjected to base white noise excitation under various damage conditions as damage data, and directly uses it as the input data of the model; divides and labels the data, and randomly shuffles them to form a complete data set; constructs an LCNN model for damage identification based on a convolutional neural network (CNN) and a long short-term memory network (LSTM); uses the LCNN model to train and update parameters on the training set, verifies the training effect each time on the validation set and fine-tunes the parameters, saves the model with the highest accuracy, and tests the final training effect on the test set; finally, performs damage identification on unlabeled data, and accurately identifies and locates the damage.

[0047] A tower structure damage identification method based on deep learning provided by this application includes the following steps:

[0048] Step S1: Establish a data analysis model of the tower structure, perform modal analysis, and extract the dynamic characteristic parameters of the structure; for a uniform and symmetric structure, Python tools can be used to batch generate components and quickly build a model, perform modal analysis on the structure and extract modal parameters;

[0049] Step S2: Batch extract the acceleration time history of vertical distribution measurement points when the structure is subjected to base white noise excitation under various damage conditions as damage data;

[0050] Step S3: Create a dataset, divide and label the dataset, and divide the dataset into a training set, a validation set, and a test set; for example, 80% can be used as the training set, 10% as the validation set, and 10% as the test set, and load the dataset into the deep learning model training module;

[0051] Step S4: Build an LCNN model for structural damage identification based on CNN and long short-term memory network LSTM;

[0052] Step S5: Use the training set to train the parameters of the LCNN model, use the validation set to verify each training result, and finally use the test set to test the trained model;

[0053] Step S6: Use the trained model to identify and locate the damage of unlabeled data. Specifically, a category can be defined for each location, a label can be defined for each category, the unlabeled data is input into the model, the model will calculate each sample and obtain a label, the label represents the category, and the category represents the location, so as to realize the location of the damage.

[0054] Further, in step S2, the following steps are included:

[0055] Step S21: Layer the structure and select the structural bearing plate as the target damage component;

[0056] Step S22: Reduce the stiffness of the target damage component to different degrees to simulate different degrees of damage;

[0057] Step S23: Extract the acceleration time history of the vertical distribution measurement points when the structure is excited by base white noise under each damage condition as damage data, and select the structural layer where the target damage component is located as the corresponding label data;

[0058] Step S24: Use a script tool such as a Python script to call a data analysis model to perform time history analysis for each condition and extract the acceleration time history of each measurement point under each condition as damage data, so as to quickly and batch generate the original data required for the deep learning training and prediction dataset.

[0059] Further, in step S3, the following steps are included:

[0060] Step S31: Cut the data generated in step S2 according to the specified sample shape;

[0061] Step S32: Set the damage layer of the data where each sample is located as the label of the sample.

[0062] Further, in step S4, the following steps are included:

[0063] Step S41: Build the CNN and LSTM sub-models respectively, which can be implemented using software such as PyTorch;

[0064] Step S411: Build the CNN model, use three consecutive one-dimensional convolutional layers to extract features from the data, and specify the output dimensions and convolutional kernel heights of the three convolutional layers respectively. After each convolutional layer, there is a batch normalization layer, an activation layer, and a Dropout layer in sequence. The activation layer selects the Relu function and specifies the Dropout ratio;

[0065] Step S412: Build the LSTM model, and specify the number of hidden layers, the number of output dimensions, and the Dropout ratio;

[0066] Step S413: Combine the final outputs of the two sub-models and input them into a fully connected layer. The number of input channels of the fully connected layer is the sum of the final output channels of the two sub-models, and the number of output channels of the fully connected layer is the number of final categories;

[0067] Step S42: Set parameters, calculate the training loss using the cross-entropy loss function, use the Adam optimization algorithm to optimize the model, set the learning rate, and configure the GPU parameters to accelerate the training.

[0068] Furthermore, in step S5, the following steps are included:

[0069] Step S51: Train the model, set the batch size and the number of training times for a single training, set the model to the training state, input the training set into the LCNN model. In each training round, obtain the output classification result through forward propagation, calculate and record the loss between the output value and the true value, update the model parameters through backpropagation, and record and output the training loss and training parameters;

[0070] Step S52: Validate the model, set the model to the evaluation state, input the validation set into the model trained each time after each training, calculate and record the accuracy rate in each validation. When the accuracy rate is greater than the highest accuracy rate in the past training process, save the model parameters at this training time to obtain the LCNN model; for example, 80% of the sample data and corresponding label data can be randomly selected as the training set, 10% as the validation set, and 10% as the test set, and the Dataset and Dataloader are used to load the dataset into the deep learning model training module;

[0071] Step S53: Test the model, load the model with the highest accuracy rate during the training process, set the model to the evaluation state, input the test set into the trained model, and calculate the test accuracy rate on the test set as the model performance evaluation index.

[0072] Furthermore, in step S6, the following steps are included:

[0073] Step S61: Load the trained model, set the model to the evaluation state, input the pre-prepared unlabeled data into the model, and calculate the model output results corresponding to the data, which are the layers where the damage is located.

[0074] A tower structure damage identification system based on deep learning provided by the present application includes a digital modeling module, a data acquisition module, a data partitioning and dataset creation module, a deep learning model construction module, a deep learning model training module, and a damage identification module;

[0075] The digital modeling module can use finite element analysis software to establish a digital model of the tower structure, group all components according to the equal quantity division principle, and layer the structure, and send the layering results to the data acquisition module;

[0076] The data acquisition module reduces the stiffness of each layer of components in the layering results of the digital modeling module to different degrees to simulate damage, extracts the acceleration time history of the vertical distribution measurement points of the tower structure under the excitation of base white noise in each damage condition as damage data; and sends the data to the data partitioning and dataset creation module;

[0077] The data partitioning and dataset creation module saves the data collected by the data acquisition module into the dataset, slices the time history data in the time dimension to generate more samples, takes the layer number where the damage of the condition to which the sample data belongs as its label, randomly selects the sample data and the corresponding label data as the training set, the validation set, and the test set according to a certain ratio. For example, 80% can be used as the training set, 10% as the validation set, and 10% as the test set, and the dataset can be loaded into the deep learning model training module using Dataset and Dataloader;

[0078] The deep learning construction module uses the dual-channel mode of LSTM and CNN to construct the LCNN model for damage identification, and sends the constructed model to the deep learning training module;

[0079] The deep learning training module uses the constructed LCNN model to train the model on the training set to update the parameters, validates the training results each time on the validation set to fine-tune the parameters, tests the finally trained model on the test set, and sends the finally trained model to the damage identification module;

[0080] The damage identification module uses the trained LCNN model to identify the test data with unknown labels and outputs the location where the damage is located.

[0081] This application collects the acceleration time histories of each measurement point in the vertical distribution under various working conditions in the numerical model, uses the original data as input data, and constructs a complete data set for damage identification algorithm training, verification, and testing through the collected working conditions, covering various possible damage situations. This application uses a deep learning model to automatically find features from the data and perform damage identification based on these features. It updates the model parameters through the training set, corrects the training direction each time through the validation set, tests the final training effect of the model through the test set, and performs unknown damage identification through unlabeled data, accurately, effectively, and quickly learning the damage characteristics of the tower structure, so as to achieve accurate identification and positioning of damage.

[0082] Although the embodiments disclosed in this application are as above, the content described is only an embodiment adopted for the convenience of understanding this application and is not used to limit this application. Any person skilled in the art within the technical field to which this application pertains may make any modifications and changes in the form of implementation and details without departing from the spirit and scope disclosed in this application. However, the scope of patent protection of this application shall still be subject to the scope defined by the appended claims.

Claims

1. A tower structure damage identification method based on deep learning, characterized in that: The method comprises the following steps: Step S1: Establish a data analysis model of the tower structure, perform modal analysis, and extract dynamic characteristic parameters of the structure; Step S2: Batch extract the acceleration time history of vertically distributed measuring points when the structure is excited by base white noise under each damage condition as damage data; Step S3: Create a data set, divide and mark the data set, and divide the data set into a training set, a validation set, and a test set; Step S4: Building an LCNN model for structural damage identification based on convolutional neural network (CNN) and long short-term memory (LSTM); Step S5: Use the training set to train the parameters of the LCNN model, use the validation set to verify each training result, and finally use the test set to test the trained model; Step S6: Use the trained model to identify and locate damage on unlabeled data.

2. The tower structure damage identification method according to claim 1, characterized in that: In step S2, the following steps are included: Step S21: stratify the structure and select the structural bearing plate as the target damaged component; Step S22: performing different degrees of stiffness reduction on the target damaged component to simulate different degrees of damage; Step S23: extracting the acceleration time history of the vertically distributed measuring points when the structure is excited by the base white noise under each damage condition as damage data, and selecting the structural layer number where the target damaged component is located as the corresponding label data; Step S24: Using a script tool to call a data analysis model to quickly generate data in batches.

3. The tower structure damage identification method according to claim 1, characterized in that: In step S3, the following steps are included: Step S31: segmenting the data generated in step S2 according to the specified sample shape; Step S32: setting the damage layer of the data where each sample is located as the label of the sample.

4. The tower structure damage identification method according to claim 1, characterized in that: In step S4, the following steps are included: Step S41: build CNN and LSTM sub-models respectively; Step S42: Set parameters, use the cross entropy damage function to calculate the training loss, use the Adam optimization algorithm to optimize the model, set the learning rate, and configure the GPU parameters to accelerate training.

5. The tower structure damage identification method according to claim 4, characterized in that: In step S41, the following steps are included: Step S411: Build a CNN model, use three consecutive one-dimensional convolutional layers to extract features from the data, specify the output dimensions and convolution kernel heights of the three convolutional layers respectively, where each convolutional layer is followed by a batch normalization layer, an activation layer, and a Dropout layer, the activation layer selects the Relu function, and specifies the Dropout ratio; Step S412: Build an LSTM model, specify the number of hidden layers, the number of output dimensions, and the Dropout ratio; Step S413: Combine the final outputs of the two sub-models and input them into a fully connected layer. The number of input channels of the fully connected layer is the sum of the final output channels of the two sub-models, and the number of output channels of the fully connected layer is the final number of categories.

6. The tower structure damage identification method according to claim 1, characterized in that: In step S5, the following steps are included: Step S51: training the model, setting the batch size and number of training times for a single training, setting the model to a training state, inputting the training set into the LCNN model, obtaining the output classification result through forward propagation in each training round, calculating and recording the loss between the output value and the true value, updating the model parameters through back propagation, and recording and outputting the training loss and training parameters; Step S52: Verify the model, set the model to the evaluation state, input the verification set into the trained model after each training, calculate and record the accuracy rate in each verification, and when the accuracy rate is greater than the highest accuracy rate in the past training process, save the model parameters under the training times to obtain the LCNN model; Step S53: Test the model, load the model with the highest accuracy during the training process, set the model to the evaluation state, input the test set into the trained model, and calculate the test accuracy on the test set as the model performance evaluation indicator.

7. The tower structure damage identification method according to claim 1, characterized in that: In step S6, the following steps are included: Step S61: load the trained model, set the model to evaluation state, input the pre-prepared unlabeled data into the model, and calculate the model output result corresponding to the data, which is the layer where the damage is located.

8. A tower structure damage identification method based on deep learning, characterized in that: It includes digital modeling module, data acquisition module, data partitioning and data set creation module, deep learning model building module, deep learning model training module and damage identification module; The digital modeling module establishes a digital model of the tower structure, groups the components and layers the structure, and sends the layering results to the data acquisition module; The data acquisition module performs different degrees of stiffness reduction on each layer of components obtained by the digital modeling module to simulate damage, extracts the acceleration time history of vertically distributed measuring points when the tower structure is excited by base white noise under various damage conditions as damage data, and sends the data to the data division and marking and data set creation module; The data partitioning and data set creation module saves the data collected by the data collection module into a data set, divides the time history data in the time dimension to generate more samples, takes the layer number of the working condition damage to which the sample data belongs as its label, randomly selects the sample data and the corresponding label data as a training set, a validation set and a test set, and loads the data set into the deep learning model training module; The deep learning building module uses the LSTM and CNN dual-channel mode to build an LCNN model for damage identification, and sends the built model to the deep learning training module; The deep learning training module uses the built LCNN model to train the model on the training set to update the parameters, verifies each training result on the validation set to adjust the parameters, tests the final trained model on the test set, and sends the final trained model to the damage identification module; The damage identification module uses the trained LCNN model to identify the test data with unknown labels and outputs the location of the damage.

Images