Shield segment support crack disease impact elastic wave detection method and device based on KNN-BiLSTM

Abstract

The invention discloses a shield segment support crack disease impact elastic wave detection method and device based on KNN-BiLSTM. The detection method specifically utilizes impact elastic wave equipment to test a shield segment support, collects the frequency, wave velocity, amplitude, frequency and curve change characteristics of Rayleigh waves propagated from the surface of a structure, and uses a network structure based on KNN-BidirectionalLSTM to carry out data processing on a multi-element elastic wave characteristic and lining disease characteristic mapping model. And obtaining the shield segment support crack depth and water permeability detection result at the detection position. According to the method, the accuracy of crack depth detection is remarkably improved, the over-fitting phenomenon is effectively relieved, the recognition accuracy and speed of the model are improved, meanwhile, the crack depth prediction result can be obtained, the crack water permeability type can be detected, and the method has important practical value and popularization significance.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent non-destructive testing, and particularly relates to a method and device for detecting impact elastic waves of shield segment support crack diseases based on KNN-BiLSTM. Background Art

[0002] Cracks are common diseases in tunnel concrete structures, and they will continue to develop during the construction and operation processes with temperature, humidity changes, and load deformations. If crack prevention and control are not carried out in a timely manner, more serious disease problems such as water seepage, corrosion, and spalling will be caused, thereby endangering the structural stability of the tunnel. Therefore, how to accurately detect the depth, state, and development trend of cracks without damaging the original structure has become a technical problem that urgently needs to be solved in the field of engineering quality inspection.

[0003] At present, there are various common non-destructive testing methods, such as ultrasonic method, acoustic wave method, impact elastic wave method, infrared thermal imaging, and penetrant testing. However, these methods have their respective limitations. The ultrasonic method is generally only applicable to the testing of relatively shallow cracks; the acoustic wave method is only applicable to in-hole testing, and its application range is limited; although infrared thermal imaging can detect surface defects, its ability to detect crack depth is insufficient; penetrant testing is mainly used to detect the permeability of materials, and it is difficult to accurately reflect the detailed conditions of cracks.

[0004] As an effective non-destructive testing method, the impact elastic wave method has low signal frequency but high energy and good spectral response performance, and is applicable to the testing of relatively deep cracks. However, the technical requirements of the impact elastic wave method are relatively high, and professional operators and complex data analysis are required. Traditional data analysis methods often rely on manual experience, and the calculation methods cannot meet the requirement of using multi-wave characteristics to reflect fracture characteristics, and it is difficult to ensure the accuracy and reliability of the detection results.

[0005] With the development of deep learning technology, using machine learning to preprocess and extract features from data can significantly reduce the risks of overfitting and underfitting and improve the generalization ability of the model. In the field of signal processing, modeling the information lost during the compression and scaling processes helps to restore the signal with high fidelity, thereby further improving the detection accuracy and precision. However, how to combine deep learning technology with the impact elastic wave method to achieve intelligent detection of crack depth and water permeability is still a technical problem that urgently needs to be solved at present.

[0006] In summary, the existing non-destructive testing methods have many deficiencies in the detection of crack depth and water permeability, and there is an urgent need for a more efficient, accurate, and reliable detection method to solve this technical problem. Summary of the Invention

[0007] To solve the above problems, the present invention specifically establishes an intelligent detection model for crack depth based on the impact elastic wave method. The impact elastic wave method is less affected by the environment, has high test efficiency and accuracy, and the intelligent detection method can reduce human errors. This method greatly improves the efficiency of the appraisal work and saves a large amount of human and financial resources.

[0008] The specific technical solution of the present invention is as follows: A method for detecting the crack depth and water permeability of shield segment support by impact elastic wave based on KNN-BiLSTM uses impact elastic wave equipment to test the shield segment support, collects the frequency, wave velocity, amplitude, frequency and curve change characteristics of Rayleigh waves propagating from the structure surface, and uses a network structure based on KNN-BidirectionalLSTM for data processing of the mapping model between multi-source elastic wave characteristics and lining disease characteristics to obtain the detection results of the crack depth and water permeability of the shield segment support; The construction method of the mapping model based on the KNN-BidirectionalLSTM network structure for multi-source elastic wave characteristics and lining disease characteristics is as follows: establish a data set, which includes the excitation source knocking contact time, the time difference between the start of the excitation source knocking and the reception of the signal at the receiving point, the Rayleigh wave velocity of the elastic shock wave, the offset distance, the frequency, the attenuation coefficient, the water-containing characteristics and the crack depth. Use the data set to train the mapping model based on the KNN-BidirectionalLSTM network structure for multi-source elastic wave characteristics and lining disease characteristics, and optimize it through the Hyperband algorithm. After reaching the number of iterations, obtain the mapping model based on the KNN-BidirectionalLSTM network structure for multi-source elastic wave characteristics and lining disease characteristics with the optimal weight parameters.

[0009] The construction method of the mapping model based on the KNN-BidirectionalLSTM network structure for multi-source elastic wave characteristics and lining disease characteristics includes the following steps: a) Data collection: Use impact elastic wave equipment to test the shield segment support model, detect the structure state by collecting the frequency, wave velocity, amplitude, frequency and curve change characteristics of Rayleigh waves propagating from the structure surface, and analyze the data detected by the detector and information such as the crack type, crack depth, average crack width, crack penetration, crack seepage characteristics and crack fractal dimension of the test model; b) Data set construction: Establish a data set based on the collected data. The data set includes the excitation source knocking contact time, the time difference between the start of the excitation source knocking and the reception of the signal at the receiving point, the Rayleigh wave velocity of the elastic shock wave, the offset distance, the frequency, the attenuation coefficient, the water-containing characteristics and the crack depth; divide the data set into a training set, a validation set and a test set according to a ratio; c) Signal preprocessing: Extract features from the collected data, and use high-pass filtering to remove noise, fill in missing values, and standardize the data.

[0010] d) Model construction: Construct a KNN-Bidirectional LSTM-Hyperband training model, where feature selection is performed by the KNN algorithm, the Bidirectional LSTM is used to capture the forward and backward dependencies of time series, and the Hyperband algorithm is used for hyperparameter optimization to reduce the model complexity and quickly find the optimal parameters; the preprocessed data is input into the constructed model to train the crack depth and water seepage information at the detection location. e) Parameter tuning: Use the Hyperband algorithm to tune the model parameters to improve the performance and generalization ability of the model, where each iteration is a complete training cycle of the entire dataset, used to evaluate the loss function in gradient descent and update the weights. f) Curve reconstruction: Use the improved model for training, and save the model with the optimal weight parameters after reaching the predetermined number of iterations. g) Testing and prediction: Call the model with the optimal weight parameters obtained in step f) to test the test set, and predict the crack depth and water permeability of the lining structure in the test set.

[0011] The impact elastic wave device described above is an SM98-24B Rayleigh wave instrument.

[0012] In signal preprocessing, the data extraction features include frequency and attenuation rate coefficient.

[0013] In the signal preprocessing step, high-pass filtering is used to remove noise, and missing values are filled in and the data is standardized.

[0014] In the dataset construction step, the division ratio of the training set, validation set, and test set is 8:1:1.

[0015] In the model construction step, the training model is based on the network structure of KNN-Bidirectional LSTM for the inversion and prediction of the mapping model between multi-source elastic wave features and lining disease features. Its process is input, encoding, decoding, and output, where the encoding part includes a local feature learning module and 1 Bidirectional LSTM module, and the decoding part is a fully connected layer.

[0016] In the model construction step, the model complexity is reduced by reducing the number of LSTM layers, and at the same time, the KNN algorithm is used for feature selection to improve the training efficiency and accuracy of the model.

[0017] In the parameter tuning step, the model weight update formula Wi，k+1 : where I i,k+1 represents the predicted crack width value of the (k + 1)-th data in the i-th step during the update process, A i,k+1 represents the predicted crack depth value of the (k + 1)-th data in the i-th step during the update process, Sq i,k+1 represents the predicted crack type value of the (k + 1)-th data in the i-th step during the update process, β i,k+1 represents the predicted crack water content characteristic value of the (k + 1)-th data in the i-th step during the update process.

[0018] A shield segment support crack disease impact elastic wave detection method and device based on KNN - BiLSTM, which can implement any of the above detection methods, includes a core processing unit with a network structure based on KNN - BidirectionalLSTM for mapping models of multi - source elastic wave characteristics and lining disease characteristics, a display screen, and an interface for connecting impact elastic wave devices.

[0019] The present invention has the following advantages compared with traditional detection methods: Compared with traditional detection methods, the present invention shows significant advantages in the detection of crack depth and water permeability.

[0020] First of all, the present invention can significantly improve the accuracy of detecting crack depth. By adopting the KNN - BiLSTM model and combining with the Hyperband algorithm for hyperparameter optimization, the present invention has achieved remarkable performance on the test set. Specifically, compared with the control groups using only KNN, LSTM or KNN - BiLSTM, the KNN - Bidirectional LSTM - Hyperband algorithm of the present invention has a lower root mean square error (RMSE) and a smaller mean absolute error (MAE) in crack depth prediction, and at the same time a higher coefficient of determination (R²), indicating that the prediction results of the present invention are more accurate and reliable.

[0021] Secondly, the present invention can effectively save detection time and cost. Traditional impact elastic wave method detection requires professional operators and complex data analysis processes, which are time - consuming and laborious. While the present invention reduces human error and improves detection efficiency through an intelligent detection method. At the same time, by optimizing the model structure and parameters, the present invention reduces the computational complexity and makes the detection process more efficient.

[0022] In addition, the present invention also has the ability to invert curves, and can reconstruct the detailed shape of cracks based on detection data, providing more intuitive and comprehensive detection results for engineering personnel. This function is of great significance for evaluating the development trend of cracks and formulating repair plans.

[0023] Finally, the present invention can not only predict the crack depth, but also detect the water permeability type of the crack. This function is of great significance for judging the impact of cracks on the safety of tunnel structures and can provide strong technical support for engineering decisions.

[0024] In summary, compared with traditional detection methods, the present invention has shown significant advantages in crack depth and water permeability type detection, not only improving the accuracy and efficiency of detection, but also reducing the detection cost, providing a new technical solution for the field of engineering quality detection. Brief Description of the Drawings

[0025] Figure 1 Framework diagram of the training method of the present invention.

[0026] Figure 2 Data acquisition method and device of the present invention.

[0027] Figure 3 Flowchart of the present invention.

[0028] Figure 4 Waveform diagram corresponding to the crack depth of the dataset of the present invention.

[0029] Figure 5 Result of the confusion matrix of the present invention. Detailed Description of the Invention

[0030] The present invention provides a method and device for detecting impact elastic waves of shield segment support cracks based on KNN-BiLSTM. The specific system framework is as Figure 1 and Figure 3 shown, including the following steps: Step 1, data collection: 1. Build a 1:1 model of the shield segment support structure of the subway tunnel; 2. Grind and clean the surface of the area to be detected to avoid unevenness on the surface; 3. Debug the use parameters of the detection device and select a suitable excitation source according to the requirements; 4. Start detection. During detection, pay attention to performing 3 detections in the same measurement area to avoid detection errors; 5. Data preprocessing; 6. Analyze the data detected by the detector and information such as the crack type, crack depth, average crack width, crack penetration, crack seepage characteristics, and crack fractal dimension of the test model; 7. Model training optimization.

[0031] Step 2, dataset construction: Establish relevant data sets based on the collected data; the main data features are the contact time of the excitation source knocking, the time difference (the time difference from the start of the excitation source knocking to the reception of the signal at the receiving point), the Rayleigh wave velocity of the elastic shock wave, the offset distance, the frequency, the attenuation coefficient, the water content, and the crack depth.

[0032] Step 3: Signal preprocessing; The data extraction features include frequency and attenuation rate coefficient, and there will be obvious missing phenomena in such data. Before data input, steps such as filtering to remove noise, filling missing values, and normalizing data are performed. The filtering method is high-pass filtering.

[0033] Step 4: Overall network structure: A network structure based on KNN-BidirectionalLSTM (abbreviated as KNN-BiLSTM) is proposed for the inversion and prediction of the mapping model between multi-source elastic wave features and lining disease features. Its process consists of 4 parts: input (Input), encoding (Encoder), decoding (Decoder), and output (Output). The encoding part includes a local feature learning module (capsule network) and 1 BidirectionalLSTM module, and the decoding part is a fully connected layer. Build a training model and reduce the number of LSTM layers to reduce the model complexity; after feature selection by the KNN algorithm, Bidirectional LSTM can capture the forward and backward dependencies of time series; use the Hyperband algorithm for hyperparameter optimization to quickly find the optimal parameters. Input the preprocessed data into the built model to train the crack depth and water seepage information at the detection location.

[0034] Step 5: Hyperband algorithm parameter tuning: Improve the performance and generalization ability of the model by finding the combination of optimal parameters. Use the Hyperband algorithm to find the optimal hyperparameters.

[0035] Step 6: Curve reconstruction: Use the improved model for training. After reaching the number of iterations, save the optimal model; One iteration mentioned above is a complete training cycle of the entire data set, which is used to evaluate the loss function in gradient descent and update the weights to achieve the optimal training result.

[0036] Step 7: Call the optimal weight parameters obtained in Step 5 to test the test set and predict the depth and water permeability.

[0037] Step 8: Deployment and application: Deploy the optimal model obtained above in a Raspberry Pi hardware device to improve its portability. In the Raspberry Pi, use multiple model compression methods to compress the model with the best evaluation metrics into a smaller model, and perform offline deployment on the Raspberry Pi device to achieve service publishing and interface calling functions. The device includes a memory and a processor. Example

[0038] Step 1, Data collection: 1. Build a 1:1 model of the shield lining structure of the subway tunnel. In this example, the experimental model size is 2m in length, 2m in width, and 1.5m in height. The thickness of the shield segment is 0.35 - 0.4m, and the designed concrete strength is C80.

[0039] 2. Polish and clean the surface of the area to be detected to avoid unevenness on the surface. 3. Select an impact elastic wave device (SM98 - 24B Rayleigh wave instrument) to debug the usage parameters of the device. The sampling frequency range is (0 - 2000Hz), and select a suitable excitation source according to requirements. The impact contact time of the excitation source is controlled within the range of 50 - 100 μs. 4. Start the detection. When detecting, pay attention to performing 3 detections within the same measurement area (10cm × 10cm) to avoid detection errors. 5. Data preprocessing; 6. Calculate based on the data detected by the SM98 - 24B Rayleigh wave instrument and analyze the information such as the crack type, crack depth, average crack width, crack penetration, crack seepage characteristics, and crack fractal dimension of the test model according to the detected data. The four crack types are through non - water - seeping (Ⅰ), through water - seeping (Ⅱ), non - through water - seeping (Ⅲ), and non - through non - water - seeping (Ⅳ), as shown in Table 1 specifically; 7. Model training and optimization.

[0040] Table 1 Partial experimental data of the example Step 2, Dataset construction: Build a relevant dataset based on the collected data. The above data features are mainly the impact contact time of the excitation source, time difference (the time difference from the start of the excitation source impact to the time when the receiving point receives the signal), Rayleigh wave velocity of the elastic shock wave, offset distance, frequency, attenuation coefficient, water - containing characteristics, and crack depth.

[0041] A total of 300 groups of experiments were collected in the experiment. The dataset was divided into a training set, a validation set, and a test set according to the ratio of 8:1:1; the ratio of the training set and the test set is usually 80% and 20%. In this example, three ratios of 8:1:1, 7:2:1, and 6:2:2 were used for test comparison. The test results show that the accuracy is higher when the ratio is 8:1:1.

[0042] Step 3, Signal preprocessing: Preprocess the collected surface wave data, and the filtering method is high-pass filtering.

[0043] Experimental platform: Windows 10 operating system, the graphics card is NVIDIA GeForce RTX 4080 GPU, and the processor is Intel Core i7-13700KF.

[0044] Step 4, Build a KNN-Bidirectional LSTM-Hyperband training model, and reduce the number of LSTM layers to reduce the model complexity; using Bidirectional LSTM after feature selection by the KNN algorithm can capture the forward and backward dependencies of time series; using the Hyperband algorithm for hyperparameter optimization can quickly find the optimal parameters.

[0045] Step 5, Hyperband algorithm parameter tuning: Improve the performance and generalization ability of the model by finding the combination of optimal parameters. In this paper, the Hyperband algorithm is used to find the optimal hyperparameters. One iteration mentioned above is a complete training cycle of the entire dataset, which is used to evaluate the loss function in gradient descent and update the weights to achieve the optimal training result.

[0046] Step 6, Use the improved model for training, and save the best weight parameters after reaching the number of iterations: One iteration mentioned above is a complete training cycle of the entire dataset, which is used to evaluate the loss function in gradient descent and update the weights to achieve the optimal training result. The model weight update formula W i，k+1 : where I i,k+1 represents the predicted crack width value of the (k + 1)-th data at the i-th time during the update process, A i,k+1 represents the predicted crack depth value of the (k + 1)-th data at the i-th time during the update process, Sq i,k+1 represents the predicted crack type value of the (k + 1)-th data at the i-th time during the update process, and β i,k+1 represents the predicted crack water content feature value of the (k + 1)-th data at the i-th time during the update process.

[0047] Step 7, Call the optimal weight parameters obtained in Step 6 to test the test set, and predict the crack depth and crack category of the test set.

[0048] In step 6, iterate repeatedly until the highest accuracy is achieved, save the weight parameters at this time, input the test set data into the improved model, and obtain the crack depth, that is, the probability of identifying the crack depth and the water permeability type.

[0049] Step 8, deploy and apply; as Figure 2 shown, the Raspberry Pi 5 is used as the core processing unit, the optimal model proposed in this application is quantized and compressed and burned into the Raspberry Pi 5 to implement service publishing and interface calling functions. The external interfaces of this device include two micro-HDMI, a USB-C power interface, two USB 2.0, two USB 3.0, and an Ethernet interface; the display and interaction device is a 10-inch touch screen display; the USB interface is used to connect the impact elastic wave device. Use Pyqt to write the user interface for displaying detection information.

[0050] In the experiment of this application, the control group KNN, the control group LSTM, the control group KNN-BiLSTM were selected for comparison with the KNN-Bidirectional LSTM-Hyperband algorithm proposed in the present invention. The error evaluation indexes are shown in Table 2. The R 2 value of the model proposed in this application is 0.926.

[0051] Table 2 Comparison results of model error evaluation Figure 4 is the waveform diagram corresponding to the crack depth of the data set of the present invention.

[0052] Figure 5 is the confusion matrix diagram of the present invention; where the abscissa is the true label and the ordinate is the predicted label. As shown in the figure, the accuracy of crack category Ⅰ is 0.9, the accuracy of crack category Ⅱ is 0.63, some labels are wrongly predicted as crack category Ⅲ, the accuracy of crack category Ⅲ is 0.81, and the accuracy of crack category Ⅳ is 0.76.

[0053] Through the verification of this embodiment, the shield segment support crack disease impact elastic wave detection method and device based on KNN-BiLSTM proposed in the present invention have achieved remarkable results in practical applications. Compared with traditional detection methods, the present invention not only greatly improves the accuracy of crack depth detection, but also successfully realizes the prediction of crack water permeability type, providing strong technical support for the maintenance and repair of tunnel and other concrete structures. At the same time, the present invention effectively reduces human errors and improves the detection efficiency through intelligent data processing and model training, and has important practical value and popularization significance.

Claims

1. A method and device for detecting impact elastic waves of shield segment support crack diseases based on KNN-BiLSTM, characterized in that , Using an impact elastic wave device, test the shield segment support with diseases, and collect the frequency, wave velocity, amplitude, frequency, and curve change characteristics of Rayleigh waves propagating on the structure surface as multi-source elastic wave characteristics. Use the network structure based on KNN-BidirectionalLSTM for data processing of the mapping model between multi-source elastic wave characteristics and lining disease characteristics to obtain an intelligent detection model for shield segment support crack diseases. Input the multi-source elastic wave characteristics into the model and output the prediction results of crack depth and water penetration type; The construction method of the network structure based on KNN-BidirectionalLSTM for the mapping model between multi-source elastic wave characteristics and lining disease characteristics is as follows: Establish a data set, which includes the excitation source knocking contact time, the time difference from the start of the excitation source knocking to the receiving point receiving the signal, the Rayleigh wave velocity of the elastic shock wave, the offset distance, the frequency, the attenuation coefficient, the water content characteristics, and the crack depth. Use the data set to train the network structure based on KNN-BidirectionalLSTM for the mapping model between multi-source elastic wave characteristics and lining disease characteristics, and optimize it through the Hyperband algorithm. After reaching the number of iterations, obtain the network structure based on KNN-BidirectionalLSTM for the mapping model between multi-source elastic wave characteristics and lining disease characteristics with optimal weight parameters.

2. The method according to claim 1, wherein The construction method of the network structure based on KNN-BidirectionalLSTM for the mapping model between multi-source elastic wave characteristics and lining disease characteristics includes the following steps: a) Data collection: Use an impact elastic wave device to test the shield segment support model, detect the structural state by collecting the frequency, wave velocity, amplitude, frequency, and curve change characteristics of Rayleigh waves propagating on the structure surface, and analyze according to the data detected by the detector and the crack type, crack depth, average crack width, crack penetration, crack seepage characteristics, and crack fractal dimension information of the test model; b) Data set construction: Establish a data set based on the collected data. The data set includes the excitation source knocking contact time, the time difference from the start of the excitation source knocking to the receiving point receiving the signal, the Rayleigh wave velocity of the elastic shock wave, the offset distance, the frequency, the attenuation coefficient, the water content characteristics, and the crack depth; Divide the data set into a training set, a validation set, and a test set according to a ratio; c) Signal preprocessing: Extract features from the collected data, and use a high-pass filtering method to remove noise, fill in missing values, and standardize the data; d) Model construction: Construct a KNN-Bidirectional LSTM-Hyperband training model, where feature selection is performed through the KNN algorithm, the forward and backward dependencies of the time series are captured by Bidirectional LSTM, and the Hyperband algorithm is used for hyperparameter optimization to reduce the model complexity and quickly find the optimal parameters; Input the preprocessed data into the constructed model to train the crack depth and water seepage information at the detection position; e) Parameter Tuning: The Hyperband algorithm is used to tune the model parameters to improve the model's performance and generalization ability. Each iteration is a complete training cycle for the entire dataset, used to evaluate the loss function in gradient descent and update the weights; f) Curve Reconstruction: Use the improved model for training. After reaching the predetermined number of iterations, save the model with the optimal weight parameters; g) Testing and Prediction: Call the model with the optimal weight parameters obtained in step f) to test the test set and predict the crack depth and water permeability of the lining structure in the test set.

3. The method according to claim 1, wherein The impact elastic wave device described is the SM98-24B Rayleigh wave instrument.

4. The method according to claim 2, wherein In signal preprocessing, the data extraction features include frequency, wave velocity, amplitude, frequency, as well as the characteristic frequency of curve change and the attenuation rate coefficient.

5. The method according to claim 2, characterized in that, In the signal preprocessing step, a high-pass filtering method is used to remove noise, fill in missing values, and standardize the data.

6. The method according to claim 2, characterized in that, In the dataset construction step, the division ratio of the training set, validation set, and test set is 8:1:

1.

7. The method according to claim 2, wherein In the model construction step, the training model is based on the network structure of KNN-BidirectionalLSTM for the inversion and prediction of the mapping model between multi-source elastic wave features and lining disease features. Its process is input, encoding, decoding, and output. The encoding part includes a local feature learning module and 1 BidirectionalLSTM module, and the decoding part is a fully connected layer.

8. The method according to claim 2, characterized in that, In the model construction step, reduce the number of LSTM layers to reduce the model complexity, and at the same time use the KNN algorithm for feature selection to improve the training efficiency and accuracy of the model.

9. The method according to claim 2, wherein In the parameter tuning step, the model weight update formula W i，k+1 : Among them, I i,k+1 represents the predicted crack width value of the (k + 1)-th data in the i-th update process, A i,k+1 represents the predicted crack depth value of the (k + 1)-th data in the i-th update process, Sq i,k+1 represents the predicted crack type value of the (k + 1)-th data in the i-th update process, β i,k+1 represents the predicted crack water content characteristic value of the (k + 1)-th data in the i-th update process.

10. A shield segment support crack depth and water penetration impact elastic wave detection device based on KNN-BiLSTM, which can implement the detection method described in any one of claims 1-9, is characterized in that, It includes a core processing unit with a network structure based on KNN-BidirectionalLSTM for the mapping model between multi-source elastic wave features and lining disease features, a display screen, and an interface for connecting the impact elastic wave device.

Images