Concrete filled steel tube void detection signal noise reduction method based on intelligent filtering

Abstract

The invention discloses a concrete filled steel tube void detection signal noise reduction method based on intelligent filtering, and the method comprises the steps: 1, carrying out the preprocessing of a knocking detection signal of a concrete filled steel tube structure, removing a DC component, and carrying out the normalization processing; 2, extracting time domain, frequency domain and time-frequency domain features; 3, preliminarily removing low-frequency interference and impulse noise through average filtering and median filtering; 4, using the convolutional neural network to extract deep features to realize noise reduction; and 5, fusing the preliminary filtering signal and the intelligent filtering signal through a deep neural network to generate a noise reduction signal. According to the multi-stage noise reduction method, the signal-to-noise ratio is remarkably increased, the main frequency and the time-frequency characteristics are reserved, the excellent noise suppression capability and the excellent signal integrity retention capability are shown in the complex noise environment, the accuracy and the reliability of concrete filled steel tube void detection are remarkably improved, and technical support is provided for civil engineering structure health monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of noise reduction for the detection of concrete-filled steel tube voids in the field of civil engineering, and specifically relates to a method for noise reduction of concrete-filled steel tube void detection signals based on intelligent filtering. Background Art

[0002] As a new type of composite structure, the concrete-filled steel tube structure is widely used in civil engineering fields such as bridges and buildings due to its high strength, high ductility, good seismic performance, and remarkable load-bearing capacity. Its health status is directly related to the safety and service life of the engineering structure. Most traditional structural health monitoring methods rely on destructive tests, which are not only time-consuming and laborious but may also cause certain damage to the structure. Therefore, the development of non-destructive testing methods has become an important trend. Among them, the impact method is widely used in the detection of internal defects (such as voids, cracks, and cavities) in concrete-filled steel tube structures due to its simple operation, low cost, and wide application range. The impact method reflects the health status of the structure by analyzing the acoustic characteristics in the impact signal. However, the impact signal is often mixed with various noise sources, including environmental noise, noise from the detection equipment itself, and random noise introduced during the signal acquisition process. The presence of these noises significantly reduces the signal-to-noise ratio of the signal, interferes with the feature extraction and subsequent analysis of the signal, and thus affects the accuracy of the detection results. Therefore, how to effectively denoise the impact signal has become an urgent problem to be solved.

[0003] In the field of non-destructive testing of concrete-filled steel tube structures, the existing signal denoising methods mainly include the following categories:

[0004] (1) Traditional filtering methods

[0005] Traditional filtering methods mainly include low-pass filtering, high-pass filtering, and band-pass filtering. These methods are based on the signal processing principle in the frequency domain and remove noise by filtering signals within a specific frequency range. For example:

[0006] Low-pass filtering: used to remove high-frequency noise.

[0007] High-pass filtering: used to remove low-frequency noise.

[0008] Band-pass filtering: used to retain signals within a specific frequency band while filtering out noise in other frequency bands.

[0009] Disadvantages: Traditional filtering methods cannot effectively distinguish between signals and noise. When the noise frequency band overlaps with the signal frequency band, it is easy to cause loss of signal details. In addition, the processing effect of these methods on complex noise environments (such as non-stationary noise) is poor.

[0010] (2) Time-domain filtering methods

[0011] Time-domain filtering methods directly process the time-domain characteristics of signals, including average filtering and median filtering:

[0012] Average filtering: Calculate the average value of the signal using a sliding window, which is suitable for removing random noise.

[0013] Median filtering: Take the median value of the signal through a sliding window, which can effectively suppress impulse noise.

[0014] Disadvantages: Such methods have limited noise reduction effects on non-stationary signals and are prone to over-smoothing the spike and edge parts of the signal, resulting in the loss of important signal features.

[0015] (3) Frequency-domain filtering methods

[0016] Frequency-domain filtering methods achieve noise reduction by converting the signal from the time domain to the frequency domain and performing frequency decomposition. Commonly used methods include:

[0017] Fast Fourier Transform (FFT): Filter out noise components in a specific frequency range through the frequency domain.

[0018] Wavelet transform: Decompose the signal using multi-scale analysis and remove noise from different scales.

[0019] Disadvantages: FFT has poor processing effects on non-linear and non-stationary noise. The wavelet transform requires the selection of appropriate wavelet basis functions, and different wavelet bases have a large impact on the results. In addition, the computational complexity of wavelet decomposition is relatively high, making it difficult to meet the requirements of real-time processing.

[0020] (4) Adaptive filtering methods

[0021] Adaptive filtering methods are noise reduction methods that dynamically adjust the filter parameters. For example: LMS (Least Mean Square) algorithm: Iteratively update the filter weights based on the characteristics of the signal and noise.

[0022] Disadvantages: Adaptive filtering methods are highly dependent on the input data. Especially when the noise characteristics change rapidly, the filtering effect decreases significantly. In addition, such methods require a large amount of training data to achieve good performance.

[0023] (5) Neural network-based filtering methods

[0024] In recent years, with the development of artificial intelligence, neural network-based filtering methods have gradually been applied to the field of signal noise reduction. By designing a Convolutional Neural Network (CNN) or a Deep Neural Network (DNN) to perform non-linear mapping on the signal, noise can be effectively suppressed while retaining the important features of the signal.

[0025] Disadvantages: Model training requires a large amount of labeled data, and the acquisition cost of high-quality labeled data is relatively high. The training process of neural networks is complex and requires a large amount of computing resources. When the training data is insufficient or the noise characteristics exceed the range of the training data, the generalization ability of the model is limited.

[0026] In the field of non-destructive testing of concrete-filled steel tubular structures, noise reduction of the signals generated by the tapping method is a key step to improve the testing accuracy. However, existing noise reduction techniques have obvious limitations and deficiencies in complex noise environments, mainly manifested as follows:

[0027] (1) Limited improvement in signal-to-noise ratio

[0028] Traditional filtering methods (such as low-pass filtering, high-pass filtering, etc.) are only based on simple signal separation in the frequency domain and are difficult to effectively process complex noises (such as non-linear noise and non-stationary noise), which easily leads to the loss of signal details.

[0029] Time-domain filtering methods (such as average filtering and median filtering) have limitations in suppressing random noises. Especially when the signal contains spike features or edge information, over-smoothing is likely to occur, and the key features of the signal cannot be effectively retained.

[0030] (2) Insufficient adaptability

[0031] Although adaptive filtering methods (such as the LMS algorithm) can dynamically adjust filtering parameters, they perform poorly in environments where the noise characteristics change rapidly. For non-stationary noises or complex multi-source noises, adaptive filtering methods usually have difficulty achieving precise filtering and are easily affected by parameter selection.

[0032] Existing technologies are usually designed for specific noise types and lack the ability to adapt to complex and diverse noise environments, and cannot handle scenarios where multiple noises are mixed in actual engineering.

[0033] (3) High computational complexity

[0034] Although frequency-domain filtering methods (such as fast Fourier transform and wavelet transform) can separate signals and noises from the frequency or time-frequency domain, their computational complexity is relatively high. Especially for wavelet transform, the computational requirements for wavelet basis selection and multi-scale decomposition are relatively high, making it difficult to meet the needs of real-time detection.

[0035] In engineering application scenarios, signal processing needs to balance efficiency and accuracy. While existing frequency-domain methods improve the noise reduction effect, they often sacrifice computational efficiency and are not suitable for large-scale or fast detection tasks.

[0036] (4) High dependence on labeled data

[0037] The intelligent filtering method based on neural network has certain advantages in signal denoising, but its model training process highly depends on a large amount of labeled data. Since obtaining high-quality labeled data in the non-destructive testing field requires high costs and complex experimental designs, this type of method is restricted in practical applications.

[0038] Meanwhile, model training requires a large amount of computing resources. When the training data is insufficient, the neural network model is prone to overfitting, resulting in poor generalization ability on new data and unable to meet the requirements of practical engineering applications.

[0039] (5) Insufficient retention of signal features

[0040] Existing denoising methods often focus on reducing the noise level while neglecting the integrity of key signal features. In the detection of concrete-filled steel tubular structures, the main frequency, peak characteristics, and mutation points of the signal are important bases for judging structural defects. Existing methods are prone to weakening these key features while suppressing noise, thereby affecting the accuracy and reliability of the detection results.

[0041] Especially for detection signals with obvious multi-scale and non-stationary characteristics, existing methods lack joint consideration of time-domain, frequency-domain, and time-frequency-domain features during processing, making it difficult to comprehensively retain the key information of the signal.

[0042] (6) Insufficient engineering practicability

[0043] The main difficulties in the application and popularization of current denoising technologies in practical engineering mainly include the following points:

[0044] 1. The method implementation is complex and requires a long time for parameter adjustment, making it difficult to be quickly deployed on the engineering site;

[0045] 2. The denoising algorithm has high requirements for computing devices and is not suitable for detection scenarios with limited resources;

[0046] 3. There is a lack of standardized performance evaluation indicators, and it is difficult to intuitively quantify the denoising effect.

[0047] The above problems lead to unsatisfactory application effects of existing technologies in complex engineering environments and limited actual usage scenarios.

[0048] The deficiencies of existing technologies in the denoising of concrete-filled steel tubular void detection signals are mainly reflected in the following aspects: limited improvement in signal-to-noise ratio, insufficient adaptability, high computational complexity, high dependence on labeled data, insufficient retention of signal features, and insufficient engineering practicability. These problems greatly limit the application effects of existing methods in practical engineering. There is an urgent need for a new type of denoising method to achieve more efficient and accurate processing of detection signals in complex noise environments, thereby improving the overall performance of non-destructive testing of concrete-filled steel tubular structures. Summary of the Invention

[0049] In view of the problems existing in the above-mentioned prior art, the present invention provides a noise reduction method for the detection signal of the void in concrete-filled steel tube based on intelligent filtering. The purpose is to, starting from the actual engineering requirements, for the detection of internal defects such as voids, cracks and cavities commonly found in concrete-filled steel tube structures, deeply integrate traditional filtering, frequency domain analysis and intelligent filtering technologies, and significantly improve the signal-to-noise ratio through multi-stage noise reduction processing, which can greatly improve the accuracy and reliability of defect detection, and can be widely applied to the health monitoring and non-destructive testing of civil engineering structures such as bridges and buildings. To solve the problems of limited signal-to-noise ratio improvement, insufficient adaptability, high computational complexity, high dependence on labeled data, insufficient retention of signal features and insufficient engineering practicability in the noise reduction of the detection signal of the void in concrete-filled steel tube in the prior art.

[0050] In order to achieve the above purpose, the specific scheme of the present invention is as follows:

[0051] The noise reduction method for the detection signal of the void in concrete-filled steel tube based on intelligent filtering includes the following steps:

[0052] Step 1, signal preprocessing: Preprocess the knocking detection signal of the concrete-filled steel tube structure, remove the DC component and perform normalization processing to obtain the preprocessed signal;

[0053] Step 2, feature extraction: Extract time domain features, frequency domain features and time-frequency domain features from the signal preprocessed in Step 1;

[0054] Step 3, preliminary filtering: Perform preliminary filtering on the signal after feature extraction in Step 2, and preliminarily remove low-frequency interference and impulse noise through average filtering and median filtering to obtain a preliminarily filtered signal;

[0055] Step 4, intelligent filtering: Use a convolutional neural network to perform deep noise reduction on the preliminarily filtered signal, extract deep features, and obtain an intelligently filtered signal;

[0056] Step 5, feature fusion: Fusion the preliminarily filtered signal and the intelligently filtered signal through a deep neural network to generate a final noise reduction signal.

[0057] Further, the formula for removing the DC component in Step 1 is as follows:

[0058] X 去直流 [n]=X[n]-μ X

[0059]

[0060] Where: X 去直流 [n] removes the DC component; X[n] is the original signal; μ Xis the signal mean value; N is the total number of signal sample points;

[0061] The formula for the normalization process is as follows:

[0062]

[0063] In the formula: X 归一化 [n] represents the normalized signal; max(|X 去直流 |) represents the maximum value of the absolute value of the signal.

[0064] Furthermore, the time-domain features described in step 2 include the mean value, variance, and peak value;

[0065] The formula for the mean value is as follows:

[0066]

[0067] The formula for the variance is as follows:

[0068]

[0069] The formula for the peak value is as follows:

[0070] Peak = max(X[n])

[0071] In the above formulas, X[n] is the original signal; μ X is the signal mean value; max(|X 去直流 |) represents the maximum value of the absolute value of the signal;

[0072] Furthermore, the frequency-domain features obtain the main frequency and spectral bandwidth of the signal through fast Fourier transform;

[0073] The calculation formula for the main frequency is as follows:

[0074] f0 = argmax|X(f)|

[0075] In the formula: f0 is the main frequency, and X(f) is the spectrum;

[0076] The calculation formula for the spectral bandwidth is as follows:

[0077] BW = f high -f low

[0078] In the formula: f high and f low are respectively the high-frequency and low-frequency boundaries when the signal energy in the spectrum drops to half of the maximum amplitude;

[0079] Further, the time-frequency domain features include wavelet transform coefficients. The multi-scale time-frequency characteristics of the signal are extracted using the wavelet transform coefficients, and the Daubechies wavelet is selected as the wavelet basis function.

[0080] The formula for the wavelet transform coefficients is as follows:

[0081]

[0082] In the formula: W j,k is the wavelet coefficient, ψ j,k (t) is the wavelet basis function; t represents the time variable of the signal, dt represents the integration variable, and it means calculating over the entire time axis.

[0083] Further, the formula for the average filtering in step 3 is as follows:

[0084]

[0085] In the formula: N ω is the sliding window size; X[n + k] represents the value of the signal at the time index n + k;

[0086] The formula for the median filtering is as follows:

[0087] X 中值 [n] = Median(X[n - k], …, X[n - k])

[0088] In the formula: k is the half length of the sliding window, and it is recommended that k = 2.

[0089] Further, the model structure of the convolutional neural network in step 4 includes:

[0090] Convolutional neural network input layer: It is used to input the preprocessed signal feature matrix, and the signal feature matrix includes time domain features, frequency domain features, and time-frequency domain features;

[0091] Convolutional layer: It is used to perform multi-layer convolution operations on the signal feature matrix to extract the deep features of the signal. The convolutional layer includes at least two layers of convolution operations, and each layer of convolution operation uses multiple convolutional kernels for feature extraction;

[0092] Pooling layer: It is used to perform max pooling operations on the feature maps output by the convolutional layer to compress the feature size and reduce the computational complexity;

[0093] Fully connected layer: It is used to flatten the pooled features and output the final denoised signal.

[0094] Further, the model structure of the deep neural network in step 5 includes:

[0095] Input layer of the deep neural network: used for inputting the preliminary filtered signal and the intelligent filtered signal;

[0096] Hidden layer: including at least two layers of fully connected nodes, with the number of nodes in each layer being 64 nodes and 32 nodes respectively, and both using the ReLU activation function;

[0097] Output layer: used for single-node output of the final noise-reduced signal.

[0098] Advantages of the present invention

[0099] 1. The noise reduction method for the debonding detection signal of concrete-filled steel tubes based on intelligent filtering in the present invention performs preliminary filtering by combining moving average filtering and median filtering. On the basis of retaining signal details, most of the low-frequency noise and pulse noise are quickly removed; wavelet transform is used for time-frequency feature extraction, combined with the convolutional neural network (CNN) model and the deep neural network (DNN) model to process data for the time-domain, frequency-domain, and time-frequency-domain features of the signal, extract key features, remove complex noise, and through the feature fusion of the preliminary filtered signal and the intelligent filtered signal, realize the output of the final noise-reduced signal. Combined with an intelligent filtering model with strong adaptability, it can not only process non-stationary noise, but also accurately retain the key characteristics of the signal (such as the main frequency, spikes, etc.), is particularly suitable for engineering scenarios with variable noise characteristics, and aims at the problem of the dependence of the neural network-based filtering method on a large amount of labeled data. In the case of insufficient data, the training data set is expanded through data augmentation techniques (such as adding noise perturbations, diverse data generation, etc.), and fine-tuning is performed on the basis of the pre-trained model through transfer learning methods, significantly improving the noise reduction performance and generalization ability of the model in small-sample scenarios, thereby greatly reducing the data acquisition cost and enhancing the practicality.

[0100] 2. The present invention comprehensively applies a variety of intelligent filtering algorithms, including average filtering, median filtering, and convolutional neural network (CNN) filtering, to achieve comprehensive suppression of different types of noise. Compared with traditional filtering, the filtering method of the present invention can increase the signal-to-noise ratio from 20 - 25 dB of the traditional method to 35 - 40 dB, and the noise suppression effect is significantly enhanced; the retention rate of signal details is increased by more than 30%, significantly reducing the risk of loss of key features.

[0101] 3. During the noise reduction process of the present invention, Daubechies wavelet is selected as the time-frequency feature extraction tool, and the wavelet basis function is optimally selected to ensure its good time-frequency localization characteristics, suitable for detecting mutation points and non-stationary features in the signal. At the same time, the present invention finely adjusts the hyperparameters of the intelligent filtering model, including the convolutional kernel size, activation function, learning rate, iteration times, etc., achieving the best balance between the noise reduction effect and the calculation efficiency, enabling the method to meet the real-time detection requirements of actual engineering.

[0102] 4. The signal processing of the present invention is based on multi-dimensional feature fusion. The detection error rate is reduced from 15% - 20% of the traditional method to 5% - 8%. In a complex environment where the noise intensity varies in the range of ±10 dB, the fluctuation of the detection result of the present invention is less than 2%, showing extremely high adaptability and stability. The present invention combines time-domain, frequency-domain, and time-frequency-domain feature extraction with intelligent filtering, significantly improving the detection accuracy and reliability of the void signal. It solves the technical problems of most traditional detection techniques relying on single filtering techniques and simple feature extraction means, with a detection error rate of about 10% - 20%, being unable to cope with the multi-dimensional characteristics of complex signals, being easily affected by noise interference, and resulting in low detection accuracy and reliability.

[0103] 5. The present invention adopts a multi-level signal processing framework, including four steps: preprocessing, preliminary filtering, intelligent filtering, and feature fusion. The preprocessing removes low-frequency interference through high-pass filtering, quickly reduces high-amplitude noise through preliminary filtering, uses a convolutional neural network (CNN) for deep filtering to extract key features, and fuses multiple filtered signals through a deep neural network (DNN) to retain key information. This multi-level signal processing method greatly improves the noise reduction effect and ensures the integrity of the signal. It solves the technical problems of existing technologies mostly adopting single filtering steps, with a simple signal processing process and lack of pertinence, resulting in poor filtering effects in complex noise environments.

[0104] 6. The present invention introduces algorithms such as convolutional neural network (CNN) and deep neural network (DNN). During the signal processing process, it can automatically optimize the filtering parameters. The algorithm automatically adjusts the filtering weights according to the input signal to adapt to different types of noise environments. It can improve robustness. In the face of non-linear noise interference, the intelligent filtering method of the present invention shows high robustness, and its filtering performance in complex environments is more than 50% better than that of traditional methods. It solves the problems of traditional methods relying on manually set parameters and empirical rules, lacking flexibility and adaptive ability, and being difficult to cope with complex and changeable noise environments.

[0105] 7. After the filtering of the present invention is completed, it comprehensively utilizes time-domain, frequency-domain, and time-frequency-domain feature information, and fuses multiple filtered signals through a deep neural network (DNN). The signal noise residue rate after multi-feature fusion is lower than 5%, and the noise reduction effect is significantly improved, far better than 15% - 20% of traditional methods. Multi-feature fusion retains the key detail information of the original signal, enhances the integrity of the signal, and provides input data with higher accuracy for void detection. It solves the technical problems of traditional filtering techniques usually relying on a single feature for signal processing, such as only using time-domain features, being difficult to comprehensively reflect the essential features of the signal, and resulting in limited detection effects.

[0106] 8. The present invention evaluates the noise reduction effect and signal fidelity through various verification means. Through the comparison of the signal-to-noise ratio (SNR), the results show that the SNR of the noise reduction of the present invention has increased by more than 50%; through spectral analysis, the significant suppression of noise in the frequency domain is verified; through wavelet transform analysis, wavelet reconstruction is performed on the signal after noise reduction, showing that the retention rate of signal details is as high as more than 90%. It solves the technical problem that traditional methods mostly evaluate the noise reduction effect through a single verification means (such as visual observation or simple numerical calculation), lacking comprehensiveness.

[0107] In summary, the present invention has outstanding technical innovation in the comprehensive application of various intelligent filtering technologies, multi-level signal processing, intelligent algorithms, and multi-feature fusion. The experimental and actual verification results show that the present invention significantly improves the signal noise reduction effect and detection accuracy of the debonding detection of concrete-filled steel tubes, provides strong technical support for structural safety assessment, and demonstrates excellent adaptability and robustness in complex noise environments. Brief Description of the Drawings

[0108] Figure 1 It is a flowchart of the signal noise reduction method for debonding detection of concrete-filled steel tubes based on intelligent filtering according to the present invention. Detailed Embodiments

[0109] The present invention will be further explained and described below in conjunction with the drawings and specific embodiments. It should be noted that the specific embodiments do not limit the scope of the rights of the present invention.

[0110] As Figure 1 shown, the present specific embodiment provides a signal noise reduction method for debonding detection of concrete-filled steel tubes based on intelligent filtering, including the following steps:

[0111] Step 1, signal preprocessing: Preprocess the knocking detection signal of the concrete-filled steel tube structure by a high-sensitivity acceleration sensor, remove the DC component and perform normalization processing to obtain a preprocessed signal, so as to reduce the amplitude fluctuation of the signal and thus enhance the stability of the signal;

[0112] The function of removing the DC component is to eliminate the DC component in the signal, make its mean value zero, and avoid affecting the subsequent frequency domain analysis and filtering. The formula for removing the DC component is as follows:

[0113] X 去直流 [n] = X[n] - μ X

[0114]

[0115] In the formula: X 去 DC[n] removes the DC component; X[n] is the original signal; μ X is the signal mean value; N is the total number of signal sample points.

[0116] The role of the normalization process is to normalize the signal amplitude to a standard range, such as [0, 1] or [-1, 1], and eliminate the influence of signals of different magnitudes on the training of the filtering model. The normalized amplitude range helps to avoid numerical imbalance, improve the numerical stability of the model, and accelerate the convergence speed of neural network training.

[0117] The formula for the normalization process is as follows:

[0118]

[0119] In the formula: X 归一化 [n] represents the normalized signal; max(|X 去直流 |) represents the maximum value of the absolute value of the signal.

[0120] Step 2, feature extraction: Extract time-domain features, frequency-domain features, and time-frequency domain features from the signal preprocessed in Step 1 to comprehensively describe the signal. The role is to extract the statistical characteristics of the signal in the time domain and capture the amplitude and distribution characteristics of the signal. The time-domain features include mean, variance, and peak value;

[0121] The formula for the mean value is as follows:

[0122]

[0123] The formula for the variance is as follows:

[0124]

[0125] The formula for the peak value is as follows:

[0126] Peak = max(|X[n]|)

[0127] In the above formulas, X[n] is the original signal; μ X is the signal mean value; max(|X 去直流 |) represents the maximum value of the absolute value of the signal;

[0128] The frequency-domain features are used to obtain the main frequency and spectral bandwidth of the signal through the fast Fourier transform (FFT). The purpose is to transform the signal into the frequency domain and analyze the frequency distribution of the signal using the spectral information of the signal, so as to analyze the frequency distribution situation in the subsequent stage, especially for extracting the main frequency and spectral bandwidth.

[0129] The formula for the Fourier transform (FFT) is as follows:

[0130]

[0131] In the formula: f is the frequency, and X(f) is the spectrum.

[0132] The main frequency is the frequency with the most concentrated energy in the signal spectrum, which is obtained by finding the position of the maximum amplitude in the spectrum. The formula is as follows:

[0133] f0 = argmax|X(f)|

[0134] In the formula: f0 is the main frequency, and X(f) is the spectrum.

[0135] The spectrum bandwidth represents the width of the signal energy distribution. It is an important feature of the spectrum and is used to measure the degree of expansion of the signal spectrum. It is defined as the spectrum range near the main frequency, where the frequency range is where the signal energy drops to half of the maximum amplitude. The spectrum bandwidth is calculated by the following formula:

[0136] BW = f high -f low

[0137] In the formula: f high and f low are the high-frequency and low-frequency boundaries when the signal energy in the spectrum drops to half of the maximum amplitude, respectively.

[0138] The time-frequency domain features are used to analyze the joint time and frequency characteristics of the signal. It includes wavelet transform coefficients. The multi-scale time-frequency characteristics of the signal are extracted using the wavelet transform coefficients. Since the Daubechies wavelet has good time-frequency localization characteristics and is suitable for detecting mutation points and non-stationary features in the signal, the Daubechies wavelet is selected as the wavelet basis function.

[0139] The formula for the wavelet transform coefficient is as follows:

[0140]

[0141] In the formula: W j,k is the wavelet coefficient, ψ j,k (t) is the wavelet basis function; t represents the time variable of the signal, dt represents the integration variable, indicating that the calculation is performed over the entire time axis.

[0142] Step 3, preliminary filtering: The signal after feature extraction in Step 2 is preliminarily filtered. Since the small window can balance signal detail retention and noise suppression and is suitable for the noise reduction requirements of the signal in the tapping method, the low-frequency interference and impulse noise are preliminarily removed through average filtering and median filtering to obtain the preliminarily filtered signal;

[0143] The formula for the average filtering is as follows:

[0144]

[0145] In the formula: N ωis the sliding window size; X[n + k] represents the value of the signal at time index n + k;

[0146] The formula for the median filtering is as follows:

[0147] X 中值 [n] = Median(X[n - k], …, X[n - k])

[0148] In the formula: k is the half-length of the sliding window, and it is recommended that k = 2.

[0149] Step 4, intelligent filtering: Use a convolutional neural network (hereinafter referred to as CNN) to perform deep noise reduction on the preliminarily filtered signal, extract deep features, and obtain an intelligent filtered signal;

[0150] The main purpose of CNN is to perform data processing and feature extraction on the time domain, frequency domain, and time-frequency domain features of the input signal, so as to effectively distinguish signals and noise.

[0151] The model structure of the said CNN includes:

[0152] Convolutional neural network input layer: Used to input the preprocessed signal feature matrix, and the signal feature matrix includes time domain features, frequency domain features, and time-frequency domain features; the size of the signal feature matrix is N×3, where "N" is the number of signal samples, and "3" represents the feature dimensions of the time domain feature, frequency domain feature, and time-frequency domain feature.

[0153] Convolutional layer: Used to perform multi-layer convolutional operations on the feature matrix to extract deep features of the signal. The convolutional layer includes at least two convolutional operations, and each convolutional operation uses multiple convolutional kernels for feature extraction; the convolutional layer in this embodiment is two convolutional operations, specifically as follows:

[0154] The first convolutional layer: Contains 32 convolutional kernels (size 3×3), and uses the ReLU activation function to extract low-level features.

[0155] The second convolutional layer: Contains 32 convolutional kernels (size 3×3), and uses the ReLU activation function to further extract deep features.

[0156] Pooling layer: Used to perform max pooling operations on the feature maps output by the convolutional layer. The pooling size is 2×2, which is used to reduce the feature dimensions, compress the feature size, and at the same time reduce the model calculation complexity;

[0157] Fully connected layer: Used to flatten the pooled features and output the final denoised signal.

[0158] The source of the said CNN dataset:

[0159] 1. The training data of CNN:

[0160] The original data is sourced from the signal acquisition of the debonding detection experiment on concrete-filled steel tubular structures. The percussion detection signals include the healthy state signals and defect state signals generated by percussion.

[0161] The data acquisition devices are high-sensitivity acceleration sensors and vibration acquisition instruments, and the sampling frequency is 10 kHz.

[0162] 2. Types of noise in CNN:

[0163] Environmental noise: Background mechanical vibration noise and wind noise;

[0164] System noise: Electronic noise of the detection equipment;

[0165] Artificial noise: Random interference during the signal acquisition process.

[0166] 1000 groups of signals were collected for each group of experiments, with healthy signals and defect signals each accounting for 50%. Artificial synthetic noise was added to the preprocessed signals to simulate different noise intensities, and the noise intensity fluctuated by 10 dB above and below the reference value.

[0167] After signal segmentation, 80% of the training data and 20% of the test data were generated.

[0168] 3. Data augmentation of CNN:

[0169] Since the convolutional neural network model requires a large amount of training data to improve the generalization ability, in the case of insufficient data, the following data augmentation techniques are adopted in this embodiment:

[0170] Adding noise perturbation: Superimposing random noise with different distributions on the signal, such as Gaussian noise, Poisson noise, etc.

[0171] Time-axis scaling: Linearly stretching and compressing the signal along the time axis.

[0172] Signal slicing: Extracting slices of a fixed length from the long-time series signal to expand the number of samples.

[0173] 4. Model training of CNN

[0174] (1) Loss function: The mean squared error (MSE) is used as the loss function, and the formula is as follows:

[0175]

[0176] Among them, Y 真实 [i] is the true signal value, and Y 预测 [i] is the predicted value.

[0177] (2) Optimization algorithm: The Adam optimizer is used, and the initial learning rate is set to 0.001.

[0178] (3) Training parameters: Batch Size: 64; Epoch: 1000; Weight initialization method: He initialization; Regularization method: Introduce Dropout(0.5) in the fully connected layer to prevent overfitting.

[0179] Fully connected layer: Used to flatten the pooled features and output the final denoised signal. To prevent the model from overfitting during training and improve the generalization ability of the model, Dropout(0.5) is introduced in the fully connected layer. The Dropout mechanism can randomly discard 50% of the neurons during training, enabling the model to learn more robust features.

[0180] 5. Model validation of CNN

[0181] Training set and test set splitting: 80% of the data is used as the training set, and 20% of the data is used as the test set.

[0182] Performance evaluation metrics: SNR improvement effect; Retention rate of spectral features of the signal before and after denoising; Mean squared error (MSE) on the test set.

[0183] Step 5, Feature fusion: Fuse the preliminarily filtered signal and the intelligent filtered signal through a deep neural network (hereinafter referred to as DNN) to generate the final denoised signal. The main purpose of DNN is to perform feature fusion on the preliminarily filtered signal and the intelligent filtered signal to further improve the denoising effect.

[0184] The model structure of the DNN includes:

[0185] Input layer of the deep neural network: Used to receive two groups of signal inputs, including: preliminarily filtered signal S 滤波 ; Intelligent filtered signal S after denoising by CNN 智能滤波 . The input data is of size N×2, where N is the total number of signal samples.

[0186] Hidden layer: Includes at least two layers of fully connected nodes, with 64 nodes and 32 nodes in each layer respectively. The ReLU activation function is used for both layers. To prevent the model from overfitting, Dropout(0.3) is introduced in the hidden layer, that is, 30% of the neurons are randomly discarded during training to improve the model's adaptability to unseen data.

[0187] Output layer: Used for single-node output of the final denoised signal. To further reduce the risk of model overfitting, Dropout(0.3) is still maintained in this layer, and the single node outputs the final denoised signal S 降噪 .

[0188] 1. Data set source of DNN:

[0189] The training data of the deep neural network comes from the fusion of the denoising output of the convolutional neural network and the preliminary filtered signal. The feature sources of the input signals are consistent and the same as those of the CNN model, including multi-dimensional features in the time domain, frequency domain, and time-frequency domain. After the signals in the test set are fused, the denoising effect is further verified.

[0190] 2. Model Training of DNN

[0191] Loss function: The mean squared error (MSE) is also used.

[0192] Optimization algorithm: Adam optimizer with an initial learning rate of 0.001.

[0193] Training parameters: Batch size: 32; Number of training epochs: 500; Weight initialization method: Xavier initialization.

[0194] 3. Data Augmentation and Transfer Learning of DNN

[0195] The data augmentation technique is the same as that of the CNN model.

[0196] Transfer learning technique is adopted: The weights of the pre-trained CNN are transferred to the DNN model to further optimize the denoising effect after fusion.

[0197] 4. Model Verification of DNN

[0198] The performance metrics are as follows:

[0199] Signal-to-noise ratio (SNR): The fusion effect is evaluated by comparing the improvement in the signal-to-noise ratio of the final output signal;

[0200] Spectral fidelity: Verify the integrity of key frequency components;

[0201] Wavelet transform analysis: Compare the time-frequency characteristics of the signals before and after denoising.

[0202] 5. Result Verification of DNN

[0203] To verify the effectiveness and reliability of the denoising method in this specific embodiment, the results of the denoising effect and signal integrity are verified through signal-to-noise ratio calculation, spectral analysis, and wavelet transform system. The verification steps are as follows:

[0204] The denoising effect is evaluated by calculating the signal-to-noise ratio (SNR) before and after denoising.

[0205] (1) The calculation formula for the signal-to-noise ratio is:

[0206]

[0207] Where: P 信号 is the true signal, P噪音 is the signal after noise reduction.

[0208] The experimental results show that the noise reduction method of this specific embodiment increases the signal-to-noise ratio from 20 - 25 dB of the traditional method to 35 - 40 dB, significantly improving the noise reduction effect.

[0209] (2) Spectrum analysis: By comparing the main frequency components of the signal spectrum before and after noise reduction, verify the retention of key frequency components. The experimental results show that the main frequency components of the signal after noise reduction are highly consistent with the original signal, and the key frequency components are effectively retained.

[0210] (3) Wavelet transform analysis: By comparing the time-frequency characteristics of the signal before and after noise reduction through wavelet transform, verify whether the important time-frequency features of the signal are retained. The experimental results show that the characteristics of the signal after noise reduction in the time-frequency domain are highly consistent with the original signal, and the key time-frequency features of the signal are effectively retained.

Claims

1. A method for noise reduction of the signal for detecting the void in concrete-filled steel tube based on intelligent filtering, characterized in that, It includes the following steps: Step 1, signal preprocessing: Preprocess the knocking detection signal of the concrete-filled steel tube structure, remove the DC component and perform normalization processing to obtain the preprocessed signal; Step 2, feature extraction: Extract time-domain features, frequency-domain features and time-frequency domain features from the signal preprocessed in Step 1; Step 3, preliminary filtering: Perform preliminary filtering on the signal after feature extraction in Step 2, and preliminarily remove low-frequency interference and impulse noise through average filtering and median filtering to obtain the preliminarily filtered signal; Step 4, intelligent filtering: Use a convolutional neural network to perform deep noise reduction on the preliminarily filtered signal, extract deep features, and obtain the intelligent filtered signal; Step 5, feature fusion: Fusion the preliminarily filtered signal and the intelligent filtered signal through a deep neural network to generate the final noise reduction signal.

2. The noise reduction method according to claim 1, wherein The formula for removing the DC component in Step 1 is as follows: X 去直流 [n]=X[n] - μ X Where: X 去直流 [n] removes the DC component; X[n] is the original signal; μ X is the signal mean value; N is the total number of signal sample points; The formula for the normalization processing is as follows: Where: X 归一化 [n] represents the normalized signal; max(|X 去直流 |) represents the maximum value of the absolute value of the signal.

3. The noise reduction method according to claim 1, wherein The time-domain features described in Step 2 include mean, variance and peak value; The formula for the mean is as follows: The formula for the variance is as follows: The formula for the peak value is as follows: Peak=max(|X[n]|) Where X[n] is the original signal; μ X is the signal mean value; max(|X 去直流 |) represents the maximum value of the absolute value of the signal.

4. The noise reduction method according to claim 1, wherein The frequency-domain features described in Step 2 obtain the main frequency and spectral bandwidth of the signal through fast Fourier transform; The calculation formula for the main frequency is as follows: f0=arg max|X(f)| Where: f0 is the main frequency, and X(f) is the spectrum; The calculation formula for the spectral bandwidth is as follows: BW = f high -f low where: f high and f low are the high-frequency and low-frequency boundaries, respectively, at which the signal energy in the frequency spectrum drops to half of the maximum amplitude.

5. The noise reduction method according to claim 1, wherein The time-frequency domain features described in Step 2 include wavelet transform coefficients. Use the wavelet transform coefficients to extract the multi-scale time-frequency characteristics of the signal, and select the Daubechies wavelet as the wavelet basis function; The formula for the wavelet transform coefficient is as follows: Where: W j,k is the wavelet coefficient, and ψ j,k (t) is the wavelet basis function; t represents the time variable of the signal, dt represents the integration variable, indicating that the calculation is performed over the entire time axis.

6. The noise reduction method according to claim 1, wherein The formula for the average filtering in Step 3 is as follows: where: N ω is the sliding window size; X[n + k] represents the value of the signal at time index n + k; The formula for the median filtering is as follows: X 中值 [n] = Median(X[n - k], …, X[n - k]) Where: k is the half length of the sliding window, and it is recommended that k = 2.

7. The noise reduction method according to claim 1, wherein The model structure of the convolutional neural network described in Step 4 includes: Convolutional neural network input layer: Used to input the preprocessed signal feature matrix, and the signal feature matrix includes time-domain features, frequency-domain features and time-frequency domain features; Convolutional layer: Used to perform multi-layer convolution operations on the signal feature matrix to extract deep features of the signal. The convolutional layer includes at least two layers of convolution operations, and each layer of convolution operation uses multiple convolutional kernels for feature extraction; Pooling layer: Used to perform max pooling operations on the feature maps output by the convolutional layer to compress the feature size and reduce the computational complexity; Fully connected layer: Used to flatten the pooled features and output the final noise reduction signal.

8. The noise reduction method according to claim 1, characterized in that, The model structure of the deep neural network described in Step 5 includes: Deep neural network input layer: Used to input the preliminarily filtered signal and the intelligent filtered signal; Hidden layer: Includes at least two layers of fully connected nodes, with 64 nodes and 32 nodes in each layer respectively, and both use the ReLU activation function; Output layer: Used for single-node output of the final noise reduction signal.

Images