A kind of steel pipe concrete void knocking acoustic detection method suitable for low signal-to-noise ratio environment

Abstract

The application provides a kind of concrete-filled steel tube void knocking acoustic detection method suitable for low signal-to-noise ratio environment, belongs to the field of bridge nondestructive testing, the method comprises: collecting actual engineering field noise signal, combined with simulated noise to produce various noise data sets, and collect clean knocking signal in a quiet environment, use clean signal to superimpose noise signal as training set, and train to obtain stacked denoising autoencoder model;Knocking and collecting concrete-filled steel tube surface audio signal in actual engineering environment, use the trained stacked denoising autoencoder model to denoise the signal to be identified;Use CNN-LSTM model to extract the time-frequency characteristics of the denoised signal, and complete the identification of void defects. This method overcomes the shortcoming that the precision of traditional knocking acoustic detection is greatly affected by environmental noise, and is suitable for concrete-filled steel tube void detection in complex and low signal-to-noise ratio environment, with high detection precision.

Description

Technical Field

[0001] This invention relates to the field of acoustic defect detection, and more particularly to an acoustic method for detecting voids in steel-concrete composite pipes by tapping, suitable for environments with low signal-to-noise ratios. Background Technology

[0002] Void detection in steel-concrete composite structures can promptly identify and repair voids that may lead to reduced load-bearing capacity, accelerated corrosion, and decreased durability, thereby preventing potential structural failures and increased maintenance costs. This is crucial for ensuring structural safety and extending service life. The percussion acoustic method is widely used for the inspection of large steel-concrete composite arch bridges and steel-concrete composite columns due to its simplicity, low cost, and lack of structural damage. However, in actual engineering inspections, environmental noise such as wind and traffic noise often interferes with the inspection process, significantly reducing the accuracy and reliability of the test data.

[0003] In existing percussion acoustic methods for detecting voids in concrete-filled steel tube structures, filter noise reduction is a commonly used signal processing method to reduce the interference of environmental noise on the detection data. However, since filter design is usually based on the assumption of stationary noise, they are difficult to effectively remove non-stationary and time-varying noise. This may not only lead to distortion of the detection data but also erroneously delete useful signal features, thus affecting the accurate identification of defects. To address this problem, scholars at home and abroad have proposed a series of methods, such as wavelet transform, empirical mode decomposition, and variational mode decomposition. Although these methods have achieved good results, they often require a large amount of signal processing knowledge to determine parameters when faced with different signals, exhibit poor adaptability, and cannot effectively achieve end-to-end noise removal, making their application in practical engineering inspections difficult. Therefore, it is necessary to design a noise reduction and void detection method for concrete-filled steel tubes based on percussion acoustics for practical engineering applications. Summary of the Invention

[0004] The purpose of this invention is to provide an acoustic detection method for voids in steel-concrete composite tubes using impact testing, suitable for low signal-to-noise ratio environments. This method addresses the technical problem of poor adaptability in existing concrete void prediction methods, which makes their application in practical engineering testing difficult. The method employs targeted noise reduction to address environmental noise encountered during actual engineering testing. Simultaneously, it utilizes CNN and LSTM to extract frequency and time domain information, fusing them to form multimodal features and constructing damage indicators, thereby enabling high-precision identification of the location of void defects in steel-concrete composite tubes.

[0005] The method includes: collecting noise signals from actual engineering sites, creating various noise datasets by combining simulated noise, and collecting clean tapping signals in a quiet environment. The clean signals are superimposed with noise signals as a training set to train a stacked convolutional denoising autoencoder model. Audio signals from the surface of steel pipe concrete are tapped and collected in the actual engineering environment. The trained stacked convolutional denoising autoencoder model is used to denoise the signal to be identified. The CNN-LSTM model is used to extract the time-frequency features of the denoised signal to complete the identification of void defects.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for noise reduction of steel-concrete composite void detection signals suitable for low signal-to-noise ratio environments using an autoencoder, characterized in that the method includes the following steps:

[0008] S1: In a quiet environment, a steel-concrete composite structure is struck at a fixed frequency and force, and the resulting audio data is collected and saved using a microphone.

[0009] S2: Automatically pick up the starting point of the tap signal. Using the Akaike information criterion, find the starting point of the tap audio. The audio of each tap is saved as a separate file.

[0010] S3: Noise signal production. In order to reproduce the noise of the actual engineering site, the actual noise on site is recorded and combined with the noise augmentation dataset of the simulated site, which is used as the noise source.

[0011] S4: Dataset preparation: Randomly add noise sources to the clean audio files extracted in S2 to obtain signal data as the training set;

[0012] S5: Network structure construction, adopting a stacked noise reduction autoencoder model. By stacking several noise reduction autoencoder layers, higher-level feature representations are learned step by step to recover clean signals from noisy inputs.

[0013] S6: Adjust hyperparameters and perform training; hyperparameter settings include: learning rate, number of training iterations, batch size, and optimizer parameters;

[0014] S7: Check whether the loss function on the validation set gradually converges, and use the signal-to-noise ratio after noise reduction and the root mean square error as two indicators to evaluate the performance of the model.

[0015] S8: If the performance of the two metrics in S7 meets the requirements, the training ends; if the requirements are not met, return to S7 to continue training; if the requirements are met, the training ends.

[0016] Furthermore, in S3, the noise sources include three types of sounds: wind noise, car noise, and bird noise. These three sounds are mixed together in different intensity ratios to create eight types of noise data, namely Noise 1 to Noise 8.

[0017] Furthermore, in S4, one or more of noise 1 to noise 8 are randomly added to the clean audio file. At the same time, a random scaling factor is applied to change the intensity of noise 1 to noise 8 and randomly add it into the sample to obtain signal data. The signal data is used as the training set, and the denoised audio data is used as the test set.

[0018] Furthermore, in S5, the detailed architecture of the convolutional denoising autoencoder is as follows: Input -> Convolutional layer 1 -> BN + PReLU -> Max pooling layer 1 -> Convolutional layer 2 -> BN + PReLU -> Max pooling layer 2 -> BN + PReLU -> Encoder LSTM layer -> Latent connection layer -> Decoder LSTM layer -> Deconvolutional layer 1 -> BN + PReLU -> Deconvolutional layer 2 -> BN + PReLU -> Deconvolutional layer 2 -> Tanh -> Output; Stacked convolutional denoising autoencoders use the output of one trained convolutional denoising autoencoder as the input of the next autoencoder.

[0019] Furthermore, in step 5, an encoder LSTM layer and a decoder LSTM layer are added to the convolutional denoising autoencoder. The encoder LSTM layer improves the model's ability to capture temporal features and enhances the denoising effect by modeling the temporal dependence of the audio in the convolutional denoising autoencoder.

[0020] Furthermore, the formulas for the two indicator evaluation models in S7 are as follows:

[0021] ;

[0022] ;

[0023] in: The original signal, For noise reduction signal, It is the energy of the original signal. It is the energy of the noise signal, that is, the difference between the original signal and the denoised signal.

[0024] Furthermore, the signal-to-noise ratio of the signal with added noise in S4 is maintained at 0-25dB.

[0025] An acoustic method for detecting voids in steel-concrete composite structures by impact, suitable for low signal-to-noise ratio environments, the method comprising the following steps:

[0026] SS1: The denoised audio signal is labeled with and without denoising, and used as the training set. A portion of the signal is used as the validation set to verify the model's capabilities.

[0027] SS2: The PSD signal is calculated using the Welch method. First, the signal is divided into multiple segments and windowed. The windowed signal segments are transformed using FFT and the power is calculated. Then, the results of all segments are averaged to obtain the PSD estimate image of the entire signal.

[0028] SS3: Feature-weighted concatenation: The frequency domain features of each audio PSD signal are extracted using a convolutional neural network, and the time domain features are extracted by directly processing each audio signal using a long short-term memory network, followed by weighted concatenation.

[0029] SS4: After the time-frequency features are concatenated, the synthesized feature vector is used for classification tasks. Data with and without data that need to be de-emptied are mixed as a test set, and the classification task is completed using the classification layer at the end of the model.

[0030] Furthermore, the PSD estimation image in SS2 is specifically represented as follows:

[0031] ;

[0032] in, K This is the total number of data segments. It is the first time after adding a window i part, For window functions, The total energy of the window function.

[0033] Furthermore, the weighted splicing in SS3 is represented as follows:

[0034] ;

[0035] This represents the frequency domain feature vector extracted by the CNN. This represents the time-domain feature vector extracted by LSTM. , ...

[0036] The present invention, by adopting the above-described technical solution, has the following beneficial effects:

[0037] This invention fully considers the noise conditions encountered in actual engineering testing. It uses self-collected noise to create a noise dataset, which is then randomly added. Compared to training with Gaussian white noise, this results in a model with better robustness. Furthermore, compared to variational mode decomposition denoising techniques, this invention does not require complex signal processing knowledge. It utilizes a stacked convolutional denoising autoencoder for adaptive denoising, and incorporates LSTM to improve the model's ability to capture the temporal features of audio signals, thus offering broader application prospects in practical engineering testing. Attached Figure Description

[0038] Figure 1 This is a flowchart summarizing the entire method;

[0039] Figure 2 A schematic diagram for finding the starting point;

[0040] Figure 3 This is a flowchart of a noise reduction method based on a stacked noise reduction autoencoder;

[0041] Figure 4 This is a flowchart of feature extraction and classification methods based on CNN and LSTM;

[0042] Figure 5 A schematic diagram illustrating the denoising principle and layer-by-layer training of a stacked denoising autoencoder;

[0043] Figure 6 This is a model architecture for a convolutional denoising autoencoder;

[0044] Figure 7 For casting steel-concrete composite specimens;

[0045] Figure 8 A diagram showing the layout of measuring points on a steel-concrete composite specimen with void defects.

[0046] Figure 9 This is a comparison chart of the time-domain curves and PSD curves before and after noise reduction;

[0047] Figure 10 For the audio before and after noise reduction Figure 7 The image shows a comparison of the prediction accuracy on the steel pipe. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.

[0049] Combination Figures 1-10As shown, this invention provides an acoustic detection method for voids in steel-concrete composite tubes, suitable for low signal-to-noise ratio environments, comprising an acoustic noise reduction method including eight steps (S1-S8) and a void prediction method, including four steps (SS1-SS4). To demonstrate an embodiment of our proposed method, we simulated void defects using low-density plastic foam and set up... Figure 8 The four void defects shown are located at different locations and have different thicknesses. They were cured at room temperature for 28 days.

[0050] In S2, after automatically picking up the starting point, the time for capturing each audio segment is guaranteed to be 0.2s, and the sampling rate is 48000hz, which is 9600 sampling points.

[0051] In S3, the recorded wind noise and vehicle noise were obtained from actual bridge site recordings, and animal noise was obtained from a publicly available dataset. Considering the actual situation, the distribution of the eight types of noise is shown in the table below.

[0052]

[0053] In S4, the noise reduction principle of the noise reduction autoencoder is as follows: Figure 5 As shown. In our method, the input audio data is supplemented with random noise, and the autoencoder training objective is to minimize the difference between the input noisy data and the autoencoder output. The loss function is expressed as... ;

[0054] in It is the original, clean data. It is the output of the autoencoder. n It represents the total number of samples.

[0055] In S4, the convolutional denoising autoencoder model architecture is as follows: Figure 6 As shown, an additional LSTM layer is added to the model for audio signals, introducing temporal dependency modeling capabilities. This allows the model to more effectively capture the temporal features of the audio signal, thereby better distinguishing and removing noise. This structure enables the denoising autoencoder to extract not only spatial features but also temporal features, which is very helpful in improving audio denoising performance.

[0056] In S4, four encoders and four decoders are stacked. Greedy training is performed layer-by-layer using these four stacked denoising autoencoders. This method trains the network layer by layer, training only one layer at a time and using the output of the previous layer as the input to the next. After training, the training data is passed through the encoder part of that layer to generate a new compressed representation, which serves as the input to the next layer. Each subsequent new layer is trained independently, using the output of the previous layer as input, until all layers have been trained. Finally, backpropagation is used to adjust the weights of all layers to optimize the overall network performance.

[0057] PReLU activation function is used at the end of each encoder and decoder. Because PReLU has learnable parameters, it can better adapt to different data characteristics, thereby improving the model's fitting ability and performance. Batch normalization (BN) technique is used to improve the training speed and stability of neural networks.

[0058] The hyperparameters set during this training process were as follows: initial learning rate 0.005, maximum number of training epochs 100, training set to validation set ratio 8:2, and batch processing times 4.

[0059] After training, check if the loss function on the validation set gradually converges. Evaluate model performance using two metrics: the signal-to-noise ratio (SNR) after denoising and the root mean square error (RMSE). Training is considered complete when the average SNR after denoising reaches above 20dB and the RMSE remains at a low level. If these conditions are not met, adjust the hyperparameters and continue training.

[0060] References for time-domain and frequency-domain effects after noise reduction Figure 9 As shown in the figure, the noise reduction effect of this method is quite good.

[0061] According to the technical solution provided by this invention, after noise reduction, a method for predicting voids in steel-concrete composite pipes based on CNN and LSTM is provided. The main purpose of using a convolutional neural network (CNN) to process the spectrogram is to utilize its powerful spatial feature extraction capability to identify frequency patterns. The LSTM model directly processes the original audio signal, extracts features, and then performs weighted concatenation. The weights are considered hyperparameters, and since voids are extremely sensitive to frequency, they are initially set to 8:2. Other hyperparameters include the number of training epochs and the initial learning rate, which are set to 50 and 0.001, respectively.

[0062] The SS2 uses a PSD image instead of a traditional spectrogram, which is more suitable for non-periodic signals such as audio.

[0063] In SS4, the Softmax function is used in the last layer of the network to perform the empty prediction.

[0064] Reference Figure 10As shown, in order to verify the effectiveness of the present invention, in a relatively noisy environment (signal-to-noise ratio less than 10 dB), the denoising method and prediction method proposed in this invention are compared with undenoised data and Bessel filter denoising method. It can be seen that the detection method proposed in this invention has high recognition accuracy, and the accuracy is significantly improved after denoising, which proves that the present invention still has strong applicability under low signal-to-noise ratio.

[0065] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for noise reduction of steel-concrete composite void detection signals suitable for low signal-to-noise ratio environments using an autoencoder, characterized in that: The method includes the following steps: Step S1: In a quiet environment, tap the steel-concrete composite structure at a fixed frequency and force, and use a microphone to collect and save the generated audio data. Step S2: Automatically pick up the starting point of the tap signal. Using the Akaike information criteria, find the starting point of the tap audio. Save the audio of each tap as a separate file. Step S3: Noise signal generation, recording actual noise on site and combining it with noise augmentation dataset from simulated site, using this as the noise source; Step S4: Dataset preparation. Randomly add noise sources to the clean audio files extracted in step S2 to obtain signal data as the training set. Step S5: Network structure construction. A stacked convolutional denoising autoencoder model is adopted. By stacking several denoising autoencoder layers, higher-level feature representations are learned step by step to recover clean signals from noisy inputs. Step S6: Adjust hyperparameters and perform training; hyperparameter settings include: learning rate, number of training iterations, batch size, and optimizer parameters; Step S7: Check whether the loss function of the validation set gradually converges, and use the signal-to-noise ratio after noise reduction and the root mean square error as two indicators to evaluate the performance of the model. Step S8: If the performance of the two metrics in step S7 meets the requirements, the training ends; if the requirements are not met, return to step S7 to continue training; if the requirements are met, the training ends. In step S3, the noise sources include wind noise, car noise, and bird noise. These three sounds are mixed together in different intensity ratios to produce eight noise data, which are noise 1 to noise 8. In step S5, the detailed architecture of the convolutional denoising autoencoder is as follows: Input -> Convolutional layer 1 -> BN + PReLU -> Max pooling layer 1 -> Convolutional layer 2 -> BN + PReLU -> Max pooling layer 2 -> BN + PReLU -> Encoder LSTM layer -> Latent connection layer -> Decoder LSTM layer -> Deconvolutional layer 1 -> BN + PReLU -> Deconvolutional layer 2 -> BN + PReLU -> Deconvolutional layer 2 -> Tanh -> Output; Stacked convolutional denoising autoencoders use the output of one trained convolutional denoising autoencoder as the input of the next autoencoder.

2. The method for noise reduction of steel-concrete composite void detection signal using an autoencoder suitable for low signal-to-noise ratio environments as described in claim 1, characterized in that: In step S4, one or more of noise 1 to noise 8 are randomly added to the clean audio file. At the same time, a random scaling factor is applied to change the intensity of noise 1 to noise 8 and randomly add it into the sample to obtain signal data. The signal data is used as the training set, and the denoised audio data is used as the test set.

3. The method for noise reduction of steel-concrete composite void detection signal using an autoencoder suitable for low signal-to-noise ratio environments as described in claim 1, characterized in that, In step S5, an encoder LSTM layer and a decoder LSTM layer are added to the convolutional denoising autoencoder.

4. The method for noise reduction of steel-concrete composite void detection signal using an autoencoder suitable for low signal-to-noise ratio environments according to claim 1, characterized in that: The formulas for the two indicator evaluation models in step S7 are as follows: ； ； in: The original signal, For noise reduction signal, It is the energy of the original signal. It is the energy of the noise signal, that is, the difference between the original signal and the denoised signal.

5. The method for noise reduction of steel-concrete composite void detection signal using an autoencoder suitable for low signal-to-noise ratio environments according to claim 1, characterized in that: In step S4, the signal-to-noise ratio of the signal with added noise is maintained at 0-25dB.

6. A method for acoustic detection of voids in steel-concrete composite structures by impact, suitable for low signal-to-noise ratio environments, characterized in that: The method includes the following steps: Step SS1: The audio signal after noise reduction by the noise reduction method described in claim 1 is labeled with "without missing space" and "without missing space" to be used as the training set, and a portion of it is used as the validation set to verify the model's capabilities. Step SS2: Calculate the PSD signal using the Welch method. First, divide the signal into multiple segments and window them. Then, transform the windowed signal segments using FFT and calculate the power. Finally, average the results of all segments to obtain the PSD estimate image of the entire signal. Step SS3: Feature weighted concatenation: The frequency domain features of each audio PSD signal are extracted using a convolutional neural network, and the time domain features are extracted by directly processing each audio signal using a long short-term memory network, followed by weighted concatenation. Step SS4: After the time-frequency features are concatenated, the synthesized feature vector is used for classification tasks. Data that needs to be de-emptied and data that are not de-emptied are mixed as a test set, and the classification task is completed using the classification layer at the end of the model.

7. The acoustic detection method for voids in steel-concrete composite pipes in low signal-to-noise ratio environments according to claim 6, characterized in that: The PSD estimation image in step SS2 is specifically represented as follows: ； in, K This is the total number of data segments. It is the first time after adding a window i part, For window functions, The total energy of the window function.

8. The acoustic detection method for voids in steel-concrete composite pipes in low signal-to-noise ratio environments according to claim 6, characterized in that: The weighted splicing in step SS3 is as follows: ； This represents the frequency domain feature vector extracted by the CNN. This represents the time-domain feature vector extracted by LSTM. , ...

Images