Anchor rod and anchor cable nondestructive detection method based on adaptive wavelet packet decomposition and deep learning

Abstract

The invention belongs to the technical field of geotechnical engineering nondestructive testing, particularly relates to an anchor rod and anchor cable nondestructive testing method based on self-adaptive wavelet packet decomposition and deep learning, and is particularly suitable for multi-defect identification and high-precision positioning under complex working conditions. According to the method, the high-resolution time-frequency analysis of the signals is realized by exciting the anchor rod and the anchor cable, collecting stress wave signals of the anchor rod and the anchor cable, preprocessing the signals and using a dynamic optimization wavelet basis and a decomposition layer number, and the defect detection precision in a complex noise environment is remarkably improved by combining a time-frequency graph convolutional neural network and multi-source feature fusion. And meanwhile, self-adaptive detection of various materials such as steel and GFRP is supported, and the method has engineering popularization and application values.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-destructive testing in geotechnical engineering, and specifically relates to a method for detecting internal defects of anchor rods and cables based on stress wave signals, which is particularly suitable for multiple defect identification and high-precision positioning under complex working conditions. Background Art

[0002] In mining engineering, anchor bolts and cables are crucial for supporting the surrounding rock of deep tunnels. They utilize pullout resistance to connect unstable and stable rock masses, limiting deformation and movement and preventing collapse and roof falls. They also apply prestress to the deep rock mass, increasing its strength and stability and controlling deformation. Therefore, nondestructive testing of anchor bolts and cables is crucial to ensure quality and support effectiveness, improve construction efficiency, and prevent accidents.

[0003] However, existing non-destructive testing technologies for anchor rods and cables suffer from significant deficiencies, such as static signal processing, weak noise immunity, and poor engineering universality. For example, a patented non-destructive testing method for anchor rod anchoring systems (CN100416269C) relies on fixed sensors and single-point signal analysis, making it difficult to accurately identify the distribution of anchoring agents and deep anchoring defects. For example, due to signal attenuation and limitations of the detection method, the missed detection rate for deep defects is as high as 30%, and the false detection rate is high in noisy environments. While neural network-based methods, such as the patented method and apparatus for non-destructive testing of the working load of non-full-length anchor rods and cables (CN114544763A), are simple to operate, they lack adaptive signal decomposition and multi-source feature fusion capabilities, making them unable to effectively distinguish between anchor position deviations and rock structure interference. In addition, other technologies in other fields, such as the patent for a tool wear prediction method based on wavelet packet decomposition and deep learning (CN111832432B), although using a similar technical framework, are oriented towards high-frequency cutting signals (1kHz-10kHz) and rely on a fixed wavelet basis (such as db4) and a preset number of decomposition layers. They are unable to adapt to the propagation characteristics of low-frequency stress waves (50Hz-8kHz) of anchor bolts and cables, and do not design dynamic noise suppression (such as the ANC algorithm) and multi-sensor collaborative analysis, resulting in a sharp decline in performance in complex mine environments (SNR≤0dB). Therefore, the development of an anchor bolt and cable detection method that integrates high-precision signal processing and intelligent decision-making is of great significance to ensuring the quality of support projects and mine safety. Summary of the Invention

[0004] The present invention aims to provide an anchor defect detection method with strong noise resistance and adaptability to changes in working conditions, effectively solving the problems of multiple defect superposition signal separation, tiny defect identification and insufficient deep positioning accuracy, so as to improve the accuracy and reliability of anchor quality detection in geotechnical engineering.

[0005] Technical Solution

[0006] The present invention provides a nondestructive testing method for anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning, comprising the following steps:

[0007] S1. Collect stress wave signals from anchor rods and cables and perform preprocessing, including detrending, bandpass filtering, normalization, and trigger alignment.

[0008] S2. Dynamically select wavelet basis functions and optimize the number of decomposition layers, perform wavelet packet decomposition on the preprocessed signal, and generate a multi-subband time-frequency matrix;

[0009] S3. Convert the time-frequency matrix into a grayscale image and input it into a pre-trained convolutional neural network to extract deep features.

[0010] S4. Integrate time domain statistical features, frequency domain energy features, and depth features, and output the defect type and location through the classifier.

[0011] As a further limitation, the pre-processing process in step S1 is as follows:

[0012] a. Detrending: Eliminate baseline drift. The formula is: Where N is the number of sliding window points, the default is N = 500;

[0013] b. Bandpass filtering: Use 4th order Butterworth filter, cutoff frequency f L =100Hz, f H =10kHz;

[0014] c. Normalization: Dynamic range is compressed to [-1, 1], the formula is Where μ is the mean and σ is the standard deviation;

[0015] d. Trigger alignment: Take the trigger signal of the impact hammer as the time zero t0 and intercept the effective signal segment t∈[t0-5ms, t0+100ms].

[0016] As a further limitation, the dynamic selection of wavelet basis functions in step S2 is specifically:

[0017] a. Candidate basis sets: Daubechies (db4-db8), Symlets (sym5-sym8), Coiflets (coif3-coif5);

[0018] b. Selection criteria: minimum reconstruction error E r =||x norm -IDWPT(DWPT(x norm ))||2, calculate the error of each basis function, select argmin(E r ).

[0019] As a further limitation, the optimization of the number of decomposition layers in step S2 is achieved by maximizing the Shannon entropy of the sub-band, iteratively calculating the entropy values ​​of the layers L=3 to L=6, and selecting the layer with the largest entropy.

[0020] As a further qualification, the entropy maximization criterion: the number of decomposition levels L is given by the Shannon entropy Determine, iteratively calculate the entropy value from L = 3 to L = 6, and select L = argmaxH (S);

[0021] Sub-band reorganization: the decomposed 2 L The sub-bands are arranged in ascending order of frequency to construct a time-frequency matrix Where N is the number of sampling points.

[0022] As a further limitation, the step S3 is specifically as follows:

[0023] A. Time-frequency diagram construction:

[0024] a. Grayscale conversion: Normalize the matrix M to [0, 255] to generate a grayscale image I∈R H×W , default H = 224, W = 224, scaled by bilinear interpolation;

[0025] b. Data augmentation: random translation (±10 pixels), rotation (±5°), and Gaussian noise (σ=0.01) are applied to I;

[0026] B. Convolutional neural network model architecture process:

[0027] a. Input layer: receives the time-frequency map scaled to 224×224 by bilinear interpolation;

[0028] b. Feature extraction layer: Use the pre-trained ResNet-3 convolutional layer and freeze the parameters of the first three layers so that they are no longer updated during training. At the same time, fine-tune the parameters of the last two layers to adapt to the characteristics of the time-frequency diagram of the anchor bolt and cable stress wave signal.

[0029] c. Classification layer: The fully connected layer outputs the defect probability P = [p0, p1, p2], which corresponds to "no defect", "shallow defect" and "deep defect" respectively.

[0030] As a further limitation, the step S4 is specifically as follows:

[0031] a. Feature fusion: Combined time domain features (kurtosis K, waveform factor F w ), frequency domain characteristics (sub-band energy ratio R e ) and CNN output P, ​​construct feature vector V=[K,F w ,R e,p0,p1,p2]; kurtosis K is used to describe the peak characteristics of the signal, and the shape factor F w Reflects the waveform shape of the signal, sub-band energy ratio R e Reflects the energy distribution of different frequency sub-bands;

[0032] b. Positioning formula: defect location in is the stress wave velocity, E is the elastic modulus of the anchor material, ρ is the material density, and Δt is determined by the time difference of the wavelet packet node energy peak.

[0033] Compared with the prior art, the present invention has the following characteristics:

[0034] 1. Improved accuracy

[0035] (1) Signal processing and feature extraction optimization

[0036] a. Detrending can eliminate baseline drift and prevent it from interfering with signal characteristics; bandpass filtering can suppress low-frequency environmental vibrations and high-frequency electromagnetic noise, making the signal purer.

[0037] b. Adaptive wavelet packet decomposition is a key step. Dynamically selecting the wavelet basis function with the smallest reconstruction error can improve the decomposition effect. Determining the number of decomposition layers based on maximizing Shannon entropy can retain the original signal information to the greatest extent.

[0038] c. The time-frequency matrix is ​​converted into a grayscale image and then input into the pre-trained convolutional neural network ResNet-34. Combining the time-domain statistical characteristics and frequency-domain energy characteristics, it can comprehensively and accurately describe the state of the anchor rods and cables, thereby improving the accuracy of defect detection.

[0039] (2) Positioning formula and error control

[0040] Positioning formula defect location (Stress wave velocity ) plays a vital role in accurately measuring stress wave velocity and the time difference Δt between wavelet packet node energy peaks, enabling accurate calculation of defect locations. In practical applications, such as anchor rod testing in Implementation Plan 1 and anchor cable testing in Implementation Plan 2, positioning errors have been as low as 0.02m and 0.08m, respectively, representing a significant improvement over traditional methods.

[0041] 2. Noise immunity

[0042] (1) Adapting wavelet packet decomposition to noise resistance

[0043] Adaptive wavelet packet decomposition not only enables high-resolution time-frequency analysis of signals but also suppresses noise. By dynamically selecting wavelet basis functions and optimizing the number of decomposition layers, the decomposed sub-band signals can better adapt to the characteristics of noisy signals, separating noise from useful signals and reducing the impact of noise on detection results.

[0044] (2) Robustness of deep learning models

[0045] The pre-trained convolutional neural network ResNet-34, after fine-tuning, effectively extracts features from the time-frequency graph of stress wave signals. The deep learning model is highly robust. When trained on noisy data, it can learn noise patterns and suppress them, accurately identifying defect characteristics in the signal. Even at a low signal-to-noise ratio (SNR) of 5dB, it maintains an accuracy rate exceeding 85%, overcoming the impact of field noise and improving adaptability in complex environments.

[0046] 3. Adaptability

[0047] (1) Adaptive wavelet packet decomposition

[0048] Adaptive wavelet packet decomposition dynamically selects appropriate wavelet basis functions and optimizes the number of decomposition layers based on the stress wave propagation characteristics of different anchor materials (such as steel and GFRP) and grouting types (such as cement slurry and resin). Different materials and grouting types can alter the frequency and amplitude of stress wave signals. This technology automatically adjusts decomposition parameters to effectively analyze different situations.

[0049] (2) Learning ability of deep learning models

[0050] By training on stress wave signal data from a large amount of anchor bolts and cables with different materials and grouting types, the system learns the signal characteristic patterns under different circumstances. When encountering a new inspection object, it can accurately determine the defect type and location based on the learned characteristics, achieving automatic adaptation to different materials and grouting types. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a system architecture diagram of the present invention;

[0052] Figure 2 This is a flow chart of the adaptive wavelet packet decomposition of the present invention;

[0053] Figure 3 This is a schematic diagram of time-frequency generation according to the present invention;

[0054] Figure 4 Improve the ResNet-34 network structure for the present invention;

[0055] Figure 5 It is the multi-source feature fusion decision module of the present invention;

[0056] Figure 6 This is a connection diagram of the non-destructive testing device for mining anchor bolts and cables used in the present invention;

[0057] Figure 7 This is a typical defect characteristic spectrum of the present invention;

[0058] Figure 8 This is a comparison diagram of the effects of the present invention and the prior art. DETAILED DESCRIPTION

[0059] The present invention provides a nondestructive testing method for anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning, which is characterized by comprising the following steps:

[0060] S1. Collect stress wave signals from anchor rods and cables and perform preprocessing, including detrending, bandpass filtering, normalization, and trigger alignment.

[0061] S2. Dynamically select wavelet basis functions and optimize the number of decomposition layers, perform wavelet packet decomposition on the preprocessed signal, and generate a multi-subband time-frequency matrix;

[0062] S3. Convert the time-frequency matrix into a grayscale image and input it into a pre-trained convolutional neural network to extract deep features.

[0063] S4. Integrate time domain statistical features, frequency domain energy features, and depth features, and output the defect type and location through the classifier.

[0064] The present invention is composed of a signal preprocessing module, an adaptive wavelet packet decomposition module, a deep learning feature extraction module and a multi-source decision module. Figure 1 : System architecture diagram, the specific steps are as follows:

[0065] S1: Signal preprocessing and synchronous trigger input

[0066] (1) Input signal:

[0067] Data source: piezoelectric accelerometer (frequency response range 0.1Hz-20kHz, sensitivity 100mV / g).

[0068] Data format: one-dimensional time domain waveform x(t), sampling rate f s ≥20kHz, duration T≥100ms.

[0069] (2) Preprocessing process:

[0070] 1. Detrending: Eliminate baseline drift. The formula is: Where N is the number of sliding window points (default N = 500), which can effectively remove the low-frequency trend components in the signal while retaining the characteristics of the stress wave signal.

[0071] 2. Bandpass filtering: Use 4th order Butterworth filter, cutoff frequency f L =100Hz, f H =10kHz, suppressing low-frequency environmental vibration and high-frequency electromagnetic noise.

[0072] 3. Normalization: Dynamic range is compressed to [-1, 1], the formula is Where μ is the mean and σ is the standard deviation. Normalization allows signals under different acquisition conditions to have a uniform scale, facilitating subsequent processing.

[0073] 4. Trigger alignment: Using the hammer's trigger signal as time zero, t0, the effective signal segment t∈[t0-5ms, t0+100ms] is captured. In actual testing, the hammer excites stress waves, and using its trigger signal as the time reference, the signal segment containing anchor bolt and cable defect information is accurately captured. This captures the complete stress wave propagation process while reducing data processing.

[0074] S2: Adaptive wavelet packet decomposition (AWPT) process, such as Figure 2 shown

[0075] (1) Dynamic wavelet basis selection:

[0076] 1. Candidate basis sets: Daubechies (db4-db8), Symlets (sym5-sym8), Coiflets (coif3-coif5).

[0077] 2. Selection criterion: minimum reconstruction error E r =||x norm -IDWPT(DWPT(x norm ))||2, calculate the error of each basis function, select argmin(E r This method of dynamically selecting wavelet basis can automatically select the most suitable basis function according to the characteristics of the signal and improve the decomposition effect.

[0078] (2) Decomposition layer optimization:

[0079] Entropy maximization criterion: The number of decomposition layers L is determined by Shannon entropy Determine, iteratively calculate the entropy values ​​from L=3 to L=6, and select L=argmaxH(S). This ensures that the decomposed sub-band signal can retain the information of the original signal to the greatest extent.

[0080] (3) Sub-band reorganization:

[0081] After decomposition, 2 L The sub-bands are arranged in ascending order of frequency to construct a time-frequency matrix Where N is the number of sampling points. After the above steps, these sub-band signals contain the characteristics of the original stress wave signal at different frequencies and time scales.

[0082] S3: Time-frequency graph generation and deep feature extraction

[0083] (1) Time-frequency diagram construction:

[0084] 1. Grayscale conversion: Normalize the matrix M to [0, 255] to generate a grayscale image I∈R H×W (Default H=224, W=224, scaled by bilinear interpolation), the time-frequency generation diagram is as follows Figure 3 shown.

[0085] 2. Data augmentation: Random translation (±10 pixels), rotation (±5°), and Gaussian noise (σ=0.01) are applied to I.

[0086] (2) CNN model architecture process, such as Figure 4 As shown:

[0087] 1. Input layer: Receives the time-frequency map, which has been bilinearly interpolated and scaled to 224×224. Bilinear interpolation and scaling can adjust the time-frequency map to a size suitable for CNN model input while preserving image features.

[0088] 2. Feature Extraction Layer: A pre-trained ResNet-34 convolutional layer was used. To fully utilize the features learned by the pre-trained model on a large-scale image dataset, the parameters of the first three convolutional layers were frozen, preventing them from being updated during training. Simultaneously, the parameters of the last two convolutional layers were fine-tuned to adapt to the characteristics of the time-frequency graph of the anchor bolt and cable stress wave signal.

[0089] 3. Classification layer: The fully connected layer outputs the defect probability P = [p0, p1, p2], corresponding to "no defect," "shallow defect," and "deep defect," respectively. The fully connected layer integrates the features extracted by the convolutional layer and outputs the corresponding defect probability based on the learned feature patterns.

[0090] S4: Multi-source decision-making and positioning calculation, multi-source feature fusion decision module such as Figure 5 As shown:

[0091] 1. Feature fusion: Combine time domain features (kurtosis K, waveform factor F w ), frequency domain characteristics (sub-band energy ratio R e ) and CNN output P, ​​construct feature vector V = [K, F w , R e , p0, p1, p2]. Kurtosis K is used to describe the peak characteristics of the signal, and the shape factor F w Reflects the waveform shape of the signal, sub-band energy ratio R e It reflects the energy distribution of different frequency sub-bands. By integrating multiple features, it can more comprehensively describe the state of anchor bolts and cables and improve the accuracy of defect detection.

[0092] 2. Positioning formula: defect location in is the stress wave velocity, E is the elastic modulus of the anchor material, and ρ is the material density. Δr is determined by the time difference of the energy peaks of the wavelet packet nodes. By analyzing the time difference of the energy peaks in the sub-band signals after wavelet packet decomposition and combining it with the stress wave velocity, the defect location can be accurately calculated.

[0093] Example 1: Anchor bolt detection

[0094] (1) Testing conditions

[0095] 1. Anchor parameters:

[0096] Material: Q345 steel (elastic modulus E = 210GPa, density ρ = 7850kg / m 3 ).

[0097] Dimensions: length L = 3m, diameter d = 25mm, grouting material is M30 cement mortar.

[0098] Defect: distance from the orifice D real =Prefabricate a cavity with a diameter of 20mm at 0.8m.

[0099] 2. Collection equipment:

[0100] High-frequency stress wave sensor: PCB 352C33 high-frequency accelerometer (frequency response range 0.5Hz-10kHz, sensitivity 100mV / g).

[0101] Data acquisition instrument: NIPXIe-4499, sampling rate f s =20kHz, resolution 24 bits.

[0102] Shock hammer trigger device: impact energy 10J±3%, trigger synchronization accuracy 0.05ms.

[0103] Tablet: Sufficient computing power, storage space, and suitable interfaces.

[0104] Cable: BNC cable.

[0105] 3. Environmental conditions

[0106] Temperature: -10℃-40℃ (simulating the actual ambient temperature in mine tunnels).

[0107] Background noise: SNR = 3dB--5dB (simulating mechanical vibration and ventilation noise)

[0108] Vibration interference: Maximum acceleration 5g (random vibration frequency 10Hz-2kHz)

[0109] (2) Signal Collection

[0110] During the connection process of the anchor bolt and cable nondestructive testing instrument (the instrument connection diagram is as follows Figure 6 The vibrating hammer generates a trigger signal at the moment of striking with its built-in trigger device. The PCB 352C33 high-frequency accelerometer used for anchor bolt detection is connected to the NIPXIe-4499 data collector via a dedicated cable, transmitting the collected analog stress wave signal to the data collector for sampling and quantification. The data collector is then connected to the tablet computer via an Ethernet interface with a network cable to transmit and save the digital signal.

[0111] (3) Signal processing flow

[0112] 1. Signal preprocessing:

[0113] Detrending: Sliding window mean filtering (window length N = 500 points, corresponding to 25ms).

[0114] Bandpass filtering: 4th-order Butterworth filter, cutoff frequency f L =100Hz, f H =8kHz.

[0115] Normalization: Dynamic range is compressed to [-1, 1], the formula is:

[0116]

[0117] 2. Adaptive Wavelet Packet Decomposition (AWPT):

[0118] Wavelet basis selection: Comparing the reconstruction errors of db6, sym6, and coif4, sym6 (minimum error E r =0.021).

[0119] Optimize the number of decomposition layers: calculate the Shannon entropy of the layer and determine L = 4 (the entropy value H = 2.37 is the maximum).

[0120] Subband reorganization: Generate 16 subbands and arrange them in ascending frequency order into a 4×4 time-frequency matrix.

[0121] 3. Time-frequency graph generation and CNN classification:

[0122] Image construction: The time-frequency matrix is ​​scaled to a 224×244 grayscale image and the pixel values ​​are normalized to [0, 255].

[0123] Data augmentation: random horizontal flipping (probability 50%) and Gaussian noise (σ = 0.02) are applied.

[0124] CNN model:

[0125] (a) Backbone network: ResNet-34 (pre-trained weights, freezing the first 3 residual blocks).

[0126] (b) Input adaptation: Modify the first-layer convolution kernel to 7×7×64.

[0127] (c) Training parameters: Adam optimizer (lr = 1e-4, batch size = 32), cross-entropy loss function.

[0128] 4. Multi-source decision-making and positioning:

[0129] Feature fusion: time domain kurtosis K = 4.21, frequency domain energy ratio R e =0.58, CNN output probability P = [0.02, 0.91, 0.07].

[0130] SVM classification: RBF kernel function (γ=0.1, C=10), output “shallow defect”.

[0131] Calculated by the formula:

[0132]

[0133] Error analysis:

[0134] Absolute error = |0.78-0.8| = 0.02m,

[0135] Positioning results and error analysis show that (reference Figure 7 Typical defect feature maps and Figure 8 Results comparison chart: The method of this invention achieved a positioning value of 0.78m (with the actual defect being 0.8m) in detecting shallow voids in tunnel steel anchors, with an error of only 2.5%. This example demonstrates the technical solution's ability to sensitively capture tiny defects and its interference resistance, meeting the stringent requirements for positioning accuracy on engineering sites.

[0136] Example 2: Anchor cable detection

[0137] (1) Preparation of conditions

[0138] 1. Anchor cable parameters:

[0139] Material: GFRP (elastic modulus E = 45GPa, density ρ = 2100kg / m 3 ).

[0140] Dimensions: length L = 15m, diameter d = 32mm, grouting material is epoxy resin.

[0141] Artificial defect: distance from the orifice D real =The entire section is fractured at 12m.

[0142] 2. Collection equipment:

[0143] High-frequency stress wave sensor: PCB 352C33 high-frequency accelerometer (frequency response range 0.5Hz-10kHz, sensitivity 100mV / g).

[0144] Data acquisition instrument: NI PXIe-4499, sampling rate f s =20kHz, resolution 24 bits.

[0145] Shock hammer trigger device: impact energy 10J±3%, trigger synchronization accuracy 0.05ms.

[0146] Tablet: has sufficient computing power, storage space, and suitable interfaces.

[0147] Cable: BNC cable.

[0148] 3. Environmental conditions

[0149] Temperature: -10℃-40℃ (simulating the actual ambient temperature in mine tunnels).

[0150] Background noise: SNR = 3dB--5dB (simulating mechanical vibration and ventilation noise)

[0151] Vibration interference: Maximum acceleration 5g (random vibration frequency 10Hz-2kHz)

[0152] (2) Signal Collection

[0153] During the connection process of the anchor rod and cable non-destructive testing instrument, the vibrating hammer generates a trigger signal at the moment of striking with the help of its built-in trigger device; the PCB 352C33 high-frequency accelerometer used for anchor rod testing is connected to the NIPXIe-4499 collector via a dedicated cable, and the collected analog stress wave signal is transmitted to the collector for sampling and quantification; the collector is then connected to the tablet computer via an Ethernet interface with a network cable to transmit and save the digital signal.

[0154] (3) Key technology differences

[0155] 1. Signal preprocessing:

[0156] The bandpass filter cutoff frequency is adjusted to f L =5kHz, f H =5kHz (adapting to the lower wave speed of GFRP).

[0157] After normalization, an adaptive noise cancellation (ANC) algorithm is added to improve the signal-to-noise ratio to SNR=8dB.

[0158] 2. Adaptive Wavelet Packet Decomposition (AWPT):

[0159] Wavelet basis selection coif4 (minimum error E r=0.018), the number of decomposition layers L = 5.

[0160] Sub-band energy focus analysis: The energy of the 28th sub-band (corresponding to 1.2-1.5kHz) accounts for 63%.

[0161] 3. Positioning algorithm correction:

[0162] GFRP wave velocity calculation:

[0163]

[0164] Time difference extraction: generalized cross-correlation (GCC-PHAT) algorithm is used, and the time difference Δt = 10.92ms.

[0165] Positioning results:

[0166] Error = 0.67%

[0167] Positioning results and error analysis show that (reference Figure 7 Typical defect feature maps and Figure 8 Effect comparison chart): In the detection of deep fractures in GFRP anchor cables, the method of the present invention achieved a positioning value of 11.92m (the actual defect was 12m) with an error of only 0.67%, verifying the method's ability to detect deep defects in low-velocity materials (such as GFRP), and maintaining high accuracy even in a strong noise environment (SNR = 3dB). This embodiment demonstrates the adaptability of the technical solution to complex materials (such as composite materials) and deep hidden defects, breaking through the limitations of traditional methods that are sensitive to wave velocity and lack noise resistance.

Claims

1. A non-destructive testing method for anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning, characterized in that: The following steps are involved: S1. Collect stress wave signals from anchor rods and cables and perform preprocessing, including detrending, bandpass filtering, normalization, and trigger alignment. S2. Dynamically select wavelet basis functions and optimize the number of decomposition layers, perform wavelet packet decomposition on the preprocessed signal, and generate a multi-subband time-frequency matrix; S 3. Convert the time-frequency matrix into a grayscale image and input it into a pre-trained convolutional neural network to extract deep features; S4. Integrate time domain statistical features, frequency domain energy features, and depth features, and output the defect type and location through the classifier.

2. The anchor bolt and cable nondestructive testing method based on adaptive wavelet packet decomposition and deep learning according to claim 1 is characterized in that: The pre-processing process described in step S1 is as follows: a. Detrending: Eliminate baseline drift. The formula is: Where N is the number of sliding window points, the default is N = 500; b. Bandpass filtering: Use 4th order Butterworth filter, cutoff frequency f L =100Hz, f H =10kHz; c. Normalization: Dynamic range is compressed to [-1, 1], the formula is Where μ is the mean and σ is the standard deviation; d. Trigger alignment: Take the trigger signal of the impact hammer as the time zero t0 and intercept the effective signal segment t∈[t0-5ms,t0+100ms].

3. The anchor bolt and cable nondestructive testing method based on adaptive wavelet packet decomposition and deep learning according to claim 1 is characterized in that: The dynamic selection of wavelet basis functions in step S2 is specifically as follows: a. Candidate basis sets: Daubechies (db4-db8), Symlets (sym5-sym8), Coiflets (coif3-coif5); b. Selection criteria: minimum reconstruction error E r =||x norm -IDWPT(DWPT(x norn ))||2, calculate the error of each basis function, select argmin(E r ).

4. The nondestructive testing method for anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning according to claim 1, characterized in that: The optimization of the number of decomposition layers in step S2 is achieved by maximizing the Shannon entropy of the sub-band, iteratively calculating the entropy values ​​of the layers L=3 to L=6, and selecting the layer with the largest entropy.

5. The nondestructive testing method for anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning according to claim 4 is characterized in that: Entropy maximization criterion: The number of decomposition layers L is determined by Shannon entropy Determine, iteratively calculate the entropy value from L = 3 to L = 6, and select L = argmaxH (S); Sub-band reorganization: the decomposed 2 L The sub-bands are arranged in ascending order of frequency to construct a time-frequency matrix Where N is the number of sampling points.

6. The method for nondestructive testing of anchor bolts and cables based on adaptive wavelet packet decomposition and deep learning according to claim 1, characterized in that: The step S3 is specifically as follows: A. Time-frequency diagram construction: a. Grayscale conversion: Normalize the matrix M to [0, 255] to generate a grayscale image I∈R H×W , default H = 224, W = 224, scaled by bilinear interpolation; b. Data augmentation: random translation (±10 pixels), rotation (±5°), and Gaussian noise (σ=0.01) are applied to I; B. Convolutional neural network model architecture process: a. Input layer: receives the time-frequency map scaled to 224×224 by bilinear interpolation; b. Feature extraction layer: Use the pre-trained ResNet-3 convolutional layer and freeze the parameters of the first three layers so that they are no longer updated during training. At the same time, fine-tune the parameters of the last two layers to adapt to the characteristics of the time-frequency diagram of the anchor bolt and cable stress wave signal. c. Classification layer: The fully connected layer outputs the defect probability P = [p0, p1, p2], which corresponds to "no defect", "shallow defect" and "deep defect" respectively.

7. The anchor bolt and cable nondestructive testing method based on adaptive wavelet packet decomposition and deep learning according to claim 1 is characterized in that: The step S4 is specifically as follows: a. Feature fusion: Combined time domain features (kurtosis K, waveform factor F w ), frequency domain characteristics (sub-band energy ratio R e ) and CNN output P, ​​construct feature vector V = [K, F w ,R e , p0, p1, p2]; kurtosis K is used to describe the peak characteristics of the signal, and the shape factor F w Reflects the waveform shape of the signal, sub-band energy ratio R e Reflects the energy distribution of different frequency sub-bands; b. Positioning formula: defect location in is the stress wave velocity, E is the elastic modulus of the anchor material, μ is the material density, and Δt is determined by the time difference of the wavelet packet node energy peak.

Images