Nondestructive testing method for flexible photovoltaic support cement grouting anchor rod pile

Abstract

The invention provides a nondestructive testing method for a cement grouting anchor rod pile of a flexible photovoltaic support, and belongs to the technical field of the cement grouting anchor rod pile of the flexible photovoltaic support, and the nondestructive testing method comprises the following steps: a preparation stage: performing insulation and anti-corrosion treatment on a reinforcing steel bar and an expanded head of the anchor rod pile; in the anchor rod construction process, a plurality of electrodes are arranged on the expansion head, the anchoring section and the free section, and positioning electrodes are arranged; resistivity measurement: after the construction of the anchor rod is completed, measuring the resistivity by using an electrical instrument, and recording measurement data; data preprocessing: performing cleaning, normalization and noise reduction processing on the measured resistivity data; feature extraction: extracting key features from the processed resistivity data, wherein the key features comprise overall features, periodic features and local features; model analysis: inputting the key features into a pre-trained nondestructive testing model, and outputting a quality vector of the anchor rod pile; the problem that irreversible structure damage is often caused by an existing detection method is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of cement grouting anchor piles of flexible photovoltaic support, and in particular, relates to a non-destructive detection method for cement grouting anchor piles of flexible photovoltaic support. Background Art

[0002] In recent years, with the rapid development of renewable energy such as photovoltaic power generation, the structural safety and reliability of photovoltaic brackets, as important infrastructure for carrying photovoltaic cell components, have received increasing attention. Photovoltaic brackets usually need to be reliably fixed to the foundation or building to withstand external loads such as wind loads and snow loads and transfer them to the foundation. Among them, the use of cement grouting anchor piles as the foundation form of photovoltaic brackets is becoming more and more common. This solution has the advantages of stable foundation and simple construction. However, due to the complex internal structure of cement grouting anchor piles and the fact that most of them are buried underground, it is difficult to directly detect and evaluate their quality status. Often, they can only rely on traditional destructive detection methods, causing irreversible structural damage. This detection method is costly and has limited detection frequency. It is difficult to detect and repair potential quality problems in a timely manner, which seriously limits the safety and service life of the photovoltaic bracket structure. Summary of the invention

[0003] In view of this, the present invention provides a non-destructive detection method for cement grouting anchor piles of a flexible photovoltaic support, which can solve the technical problem that the existing detection of cement grouting anchor piles of a flexible photovoltaic support often causes irreversible structural damage.

[0004] The present invention is achieved in that:

[0005] The present invention provides a nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support, which comprises the following steps:

[0006] S10, Preparation stage: Insulate and treat the steel bars and enlarged heads of the anchor piles to ensure the reliability of the electrode layout; At this stage, the steel bars and enlarged heads of the anchor piles need to be insulated and treated to ensure the reliability of the electrode layout. The purpose of this step is to ensure good contact between the electrode and the anchor pile during the subsequent resistivity measurement process to avoid measurement errors due to corrosion or poor insulation. Through such pretreatment, the electrical connection between the electrode and the anchor pile can be ensured to be stable, laying the foundation for subsequent resistivity measurements.

[0007] S20, electrode layout: During the anchor construction process, multiple electrodes are laid out in the enlarged head, anchoring section and free section, and the positioning electrode layout is performed; at this stage, during the anchor construction process, multiple electrodes need to be laid out in the enlarged head, anchoring section and free section respectively, and positioned. In other words, multiple measuring electrodes need to be laid out in the key parts of the anchor pile, including the enlarged head, anchoring section and free section. This distributed electrode layout can capture the resistivity characteristics of different parts and provide richer information for subsequent data analysis and feature extraction. At the same time, the specific position of each electrode needs to be recorded to provide a basis for subsequent data analysis and result interpretation.

[0008] S30, Resistivity measurement: After the anchor construction is completed, use the electrical instrument to measure the resistivity and record the measurement data; At this stage, it is necessary to use the electrical instrument to measure the resistivity and record the measurement data after the anchor construction is completed. The measurement data includes:

[0009] 1. Electrode position P i : represents the spatial coordinates of the i-th electrode.

[0010] 2. Supply current I: The current passing through the electrode pair, measured in amperes (A).

[0011] 3. Measure voltage U: Measure the voltage difference between the electrode pair in volts (V).

[0012] 4. Calculate resistivity ρ: According to Ohm's law The calculated resistivity value is in ohm-meter (Ω·m).

[0013] 5. Measurement time t: indicates the time of measurement.

[0014] 6. Ambient temperature T: The ambient temperature during measurement, in degrees Celsius (℃).

[0015] 7. Soil moisture content M: The soil moisture content at the time of measurement, expressed in volume percentage (%).

[0016] These measurement data provide the basis for subsequent data preprocessing, feature extraction and model analysis.

[0017] S40, data preprocessing: cleaning, normalizing and noise reduction of the measured resistivity data; In this stage, the measured resistivity data needs to be cleaned, normalized and noise reduced. Specifically including:

[0018] 1. Cleaning: Eliminate abnormal data points, such as obvious outliers caused by instrument failure or environmental interference.

[0019] 2. Normalization: Map data of different dimensions (such as resistivity, temperature, water content, etc.) to the same numerical range to avoid certain features dominating the model training.

[0020] S50, feature extraction: extracting key features from the processed resistivity data, including overall features, periodic features and local features;

[0021] S60, model analysis: input the key features into a pre-trained non-destructive testing model, and output the quality vector of the anchor pile.

[0022] On the basis of the above technical solution, the nondestructive testing method of cement grouting anchor piles of a flexible photovoltaic support of the present invention can also be improved as follows:

[0023] The measurement data include electrode position, power supply current, measurement voltage, calculated resistivity, measurement time, ambient temperature, and soil moisture content.

[0024] Furthermore, the overall characteristics include mean, variance, peak, skewness, and kurtosis.

[0025] Furthermore, the periodic characteristics are period-related characteristics of each type of continuous resistivity data of different time lengths, including time window characteristics, period length, period fluctuation amplitude, period phase difference, and period frequency of the continuous resistivity data of different time lengths.

[0026] Furthermore, the time window characteristics of the continuous resistivity data of different time lengths specifically refer to segmenting the resistivity data using a sliding time window and extracting statistical characteristics within each time window, including mean, standard deviation, maximum, minimum, kurtosis and skewness.

[0027] Furthermore, the local features are local abnormal features in the resistivity data, including abnormal point positions, abnormal point amplitudes, abnormal point durations, and abnormal point shapes.

[0028] Furthermore, the mass vector is specifically a multidimensional vector including the opening degree of the enlarged head, the grouting quality, the integrity of the anchoring section, the state of the free section and the deformation of the steel bar.

[0029] Furthermore, the nondestructive testing model adopts a multi-branch convolutional neural network structure, including an overall feature detection branch, a periodic feature detection branch, a local feature detection branch, a superimposed attention module and a fusion branch.

[0030] Furthermore, the overall feature detection branch is used to extract global resistivity features, the input is an overall feature vector, the output is a global feature representation, and the structure is a multi-layer fully connected network;

[0031] The periodic feature detection branch is used to identify the periodic pattern of resistivity data, the input is a periodic feature sequence, the output is a periodic feature representation, and the structure is a one-dimensional convolutional neural network;

[0032] The local feature detection branch is used to capture local abnormal features, the input is a local feature sequence, the output is a local feature representation, and the structure is an attention-enhanced convolutional neural network;

[0033] The superposition attention module is used to perform weighted fusion on the outputs of different feature branches, and its structure is a multi-head self-attention mechanism;

[0034] The fusion branch is used to integrate all feature representations and output the final result. The input is the feature representation after weighted fusion, and the output is the anchor pile mass vector. The structure is a fully connected network and a softmax layer.

[0035] Furthermore, the steps of establishing the training data set of the non-destructive testing model are specifically:

[0036] Step 11, obtaining multiple groups of cement grouting anchor piles of flexible photovoltaic brackets with a single quality defect, and using steps S10-S50 to obtain the key features of each group of anchor piles, recorded as the first feature; the quality defects specifically include incomplete opening of the enlarged head, insufficient grouting, cracks in the anchoring section, corrosion in the free section, and deformation of the steel bar;

[0037] Step 12, obtaining multiple groups of cement grouting anchor piles of flexible photovoltaic supports with at least two quality defects, and using steps S10-S50 to obtain key features of each group of anchor piles, recorded as the second feature;

[0038] Step 13, group all types of quality defects into pairs to form defect pairs, obtain each type of flexible photovoltaic support cement grouting anchor pile with defect pairs, and use steps S10-S50 to obtain the key features of each group of anchor piles, which are recorded as the third features;

[0039] Step 14: According to each defect pair, the relative influence of the two defects in the defect pair on the feature is calculated according to the corresponding third feature and the two first features;

[0040] Step 15: According to the relative influence of all defect pairs, a feature influence weight matrix is ​​established to adjust the importance of different features in the model and to initialize and constrain the attention weight of the superposition attention module;

[0041] Step 16: Establish a training data set based on the weight matrix, the first feature, the second feature, and the third feature, wherein the input of the training is the key feature, and the output of the training is the corresponding quality defect label.

[0042] Furthermore, the calculation formula involved in the feature extraction step is as follows:

[0043] 1. Overall feature extraction:

[0044] F g =[μ,σ 2 ,P max ,S k ,K];

[0045] In the formula, F g is the overall eigenvector; μ is the mean resistivity; σ 2 is the resistivity variance; P max is the resistivity peak; S k is the resistivity skewness; K is the resistivity kurtosis.

[0046] Calculation method:

[0047]

[0048] Where, N is the number of resistivity measurement points; ρ i is the resistivity value of the i-th measurement point.

[0049] 2. Periodic feature extraction:

[0050]

[0051] In the formula, F p is the period feature vector; W is the time window feature; T is the period length; A is the period fluctuation amplitude; is the periodic phase difference; f is the periodic frequency.

[0052] Calculation method:

[0053] W j =[μ j ,σ j ,max j ,min j ,K j ,S k,j ];

[0054]

[0055]

[0056] Among them, W j is the feature of the jth time window; M is the number of cycles identified; t j is the starting time of the jth cycle; ρ max,j and ρ min,j are the maximum and minimum resistivity of the jth period respectively; t peak is the peak time; tref For reference time.

[0057] 3. Local feature extraction:

[0058] F l =[P a ,A a ,D a ,S a ];

[0059] In the formula, F l is the local eigenvector; P a is the location of the abnormal point; A a is the abnormal point amplitude; D a is the duration of the abnormal point; S a is the outlier shape factor.

[0060] Calculation method:

[0061]

[0062] Among them, θ is the abnormal threshold, and the default value is 3; is the second-order derivative of resistivity, which is used to characterize the shape of the abnormal point.

[0063] Furthermore, the feature influence weight matrix calculation formula is:

[0064]

[0065] Where, W ij is the relative impact weight of defect i and defect j; F ij is the third feature including defects i and j; F i and F j are the first features containing only defects i and j respectively; ||·||2 represents the L2 norm.

[0066] Furthermore, the superposition attention module formula is:

[0067]

[0068] In the formula, α ijk is the attention weight; Q i ,K j ,V k are query, key and value matrices respectively; d k is the dimension of the key; W ij is the feature influence weight matrix.

[0069] Furthermore, the training steps of the nondestructive testing model are specifically:

[0070] Step 21: Initialize the model parameters, including the weights and biases of each branch network, and the parameters of the superimposed attention module;

[0071] Step 22: for the stacked attention module, considering the feature impact weight matrix, applying residual connection and layer normalization, and enhancing the expression ability and training stability of the model;

[0072] Step 23: Use mini-batch stochastic gradient descent method to optimize model parameters, and the loss function adopts cross entropy loss;

[0073] Step 24: Apply early stopping and regularization techniques to prevent overfitting and monitor model performance on the validation set.

[0074] Step 25: Repeat steps 23-24 until the model converges or reaches the preset number of training rounds, and save the optimal model parameters.

[0075] Through such model design and training process, we can make full use of information of different features to improve the accuracy and robustness of nondestructive testing. The output quality vector includes multiple dimensions such as the opening degree of the expansion head, grouting quality, anchor section integrity, free section state and steel bar deformation, which can provide a basis for subsequent analysis and decision-making.

[0076] Compared with the prior art, the nondestructive testing method for cement grouting anchor piles of flexible photovoltaic support provided by the present invention can comprehensively evaluate the quality status of anchor piles by laying multiple electrodes inside the anchor piles, collecting high-resolution resistivity data, and combining with advanced data analysis algorithms, so as to provide reliable protection for the safety of photovoltaic support structures. Compared with traditional destructive testing methods, this method has the following key advantages:

[0077] 1. Non-destructive: This method does not require any destructive operation on the anchor piles. It only requires the simple laying of electrodes during the construction process to obtain the required resistivity data, thus avoiding damage to the structural integrity. This not only reduces the cost of testing, but also greatly increases the frequency of testing, making it possible to monitor and diagnose the quality of anchors more timely and effectively.

[0078] 2. Comprehensiveness: This method can simultaneously capture the resistivity characteristics of each key part inside the anchor pile (such as the enlarged head, anchor section, and free section). Through an advanced multi-branch neural network model, it can achieve a comprehensive evaluation of multiple dimensions such as the enlarged head opening degree, grouting quality, anchor section integrity, free section status, and steel bar deformation, providing a basis for accurate diagnosis of the structural health status.

[0079] 3. Sensitivity: This method uses the rich information contained in the resistivity data, such as overall characteristics, periodic characteristics and local characteristics, to sensitively capture the tiny quality defects inside the anchor pile, creating conditions for timely discovery and repair of potential problems. Compared with traditional methods that can only detect obvious defects, the method of the present invention has higher detection sensitivity and fault diagnosis capabilities.

[0080] 4. Intelligence: This method uses advanced convolutional neural networks and attention mechanisms to automatically extract key features from massive resistivity data and output multi-dimensional evaluation results of anchor pile quality through an end-to-end learning process. This intelligent data analysis method greatly improves the reliability and consistency of detection and reduces dependence on professional experience.

[0081] In summary, the nondestructive testing method based on resistivity measurement proposed in the present invention makes full use of the rich information contained in the resistivity data, and realizes a comprehensive assessment of the quality status of the cement grouting anchor piles of the photovoltaic support through advanced data analysis algorithms, thereby providing strong technical support for the safety and reliability of the photovoltaic support structure, and solving the technical problem that the existing inspection of cement grouting anchor piles of flexible photovoltaic supports often causes irreversible structural damage. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A flow chart of the method provided by the present invention;

[0083] Figure 2 Schematic diagram of feature influence matrix in the embodiment. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0085] like Figure 1 As shown, it is a flow chart of a nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support provided by the present invention, and the method comprises the following steps:

[0086] S10, preparation stage: insulation and anti-corrosion treatment of the steel bars and enlarged heads of the anchor piles to ensure the reliability of electrode layout;

[0087] S20, electrode layout: during the anchor bolt construction process, multiple electrodes are laid out at the enlarged head, anchoring section and free section, and positioning electrode layout is performed;

[0088] S30, resistivity measurement: after the anchor construction is completed, use the electrical instrument to measure the resistivity and record the measurement data;

[0089] S40, data preprocessing: cleaning, normalizing and noise reduction of the measured resistivity data;

[0090] S50, feature extraction: extracting key features from the processed resistivity data, including overall features, periodic features and local features;

[0091] S60, model analysis: input key features into the pre-trained non-destructive testing model and output the quality vector of the anchor pile.

[0092] The specific implementation methods of the above steps are described in detail below:

[0093] The specific implementation method of step S10 is: first, the steel bars of the anchor pile are comprehensively inspected to ensure that there are no defects such as rust and cracks on the surface. Then, the surface of the steel bar is subjected to the first anti-corrosion treatment using epoxy resin coating, and the coating thickness is controlled at 0.2-0.3 mm. After the epoxy resin is cured, the steel bar is wrapped with polyethylene plastic film for the second time, and the film thickness is 0.1-0.15 mm. For the steel bar joints, heat shrink tubes are used for additional protection, and the length of the heat shrink tubes should cover 10 cm on both sides of the joint. The enlarged head part is treated with a special anti-corrosion coating, and the coating is selected from epoxy resin-based materials with strong alkali resistance and good bonding strength, and the coating thickness is controlled at 0.5-0.8 mm. Finally, the treated steel bar and enlarged head are tested for insulation performance to ensure that the insulation resistance is not less than 100 megohms. The purpose of this step is to provide a reliable insulation foundation for the subsequent electrode layout to prevent the electrode from directly contacting the steel bar and causing measurement errors.

[0094] The specific implementation method of step S20 is: first, according to the structural design drawing of the anchor pile, determine the specific position and size of the enlarged head, anchoring section and free section. In the enlarged head part, 16 electrodes are evenly arranged along the axial direction, and the electrode spacing is 5 cm. The electrode is made of stainless steel, with a diameter of 2 mm and a length of 5 cm. The electrode is fixed to the central steel pipe of the enlarged head by a special spring device to ensure that the electrode can be close to the cement slurry when the enlarged head is opened. In the anchoring section, 48 electrodes are arranged along the axial direction, and the electrode spacing is also 5 cm. These electrodes are made of flexible conductive materials to adapt to possible deformation of the anchoring section. In the free section, an electrode is arranged every 50 cm according to its length to monitor the overall state of the free section. All electrodes are connected to the ground data acquisition equipment through high-strength, corrosion-resistant wires. In addition, a measuring line dedicated to positioning is also arranged at the enlarged head, which contains 10 micro electrodes with a spacing of 2 cm. These electrodes are enclosed in a special electrode cable box. When the expansion head is opened, the cable is pulled out, and the opening degree of the expansion head can be judged by the number of electrodes pulled out. The purpose of this step is to build a comprehensive and accurate resistivity measurement network to provide basic data for subsequent non-destructive testing.

[0095] The specific implementation method of step S30 is: 14 days after the construction of the anchor pile is completed (until the cement slurry reaches a certain strength), the resistivity measurement is carried out. First, all electrodes are connected to a high-precision resistivity meter. The measuring instrument adopts the principle of the four-electrode method, the operating frequency range is 0.1-1000 Hz, and the measurement accuracy is better than 0.1%. During the measurement, an alternating current (frequency of 100 Hz, current intensity of 1 mA) is applied to each pair of adjacent electrodes in turn, and the voltage is measured at the same time. The resistivity is calculated according to Ohm's law. In order to improve the measurement accuracy, each measurement point is measured three times and the average value is taken. At the same time, the specific position (depth), power supply current value, measured voltage value, calculated resistivity value, measurement time, ambient temperature (accurate to 0.1 degrees Celsius) and soil moisture content (measured by time domain reflection method, accurate to 0.1%) of each measurement point are recorded. During the measurement process, a data recorder is used to automatically store all measurement data to avoid human recording errors. The purpose of this step is to obtain the detailed resistivity distribution of each part of the anchor pile and provide raw data for subsequent data analysis.

[0096] The specific implementation of step S40 is: first, detect and clean the original resistivity data for outliers. Use the 3-times standard deviation method to identify outliers, that is, mark data points that deviate from the mean value by more than 3 times the standard deviation as outliers. For the detected outliers, if it is a single isolated point, use the average value of the adjacent points before and after to replace it; if it is multiple consecutive points, use the interpolation method to correct it. Then, use the wavelet transform method to reduce the noise of the data. Select the Daubechies wavelet basis, the decomposition layer number is 4, and the soft threshold method is used for coefficient shrinkage. The threshold selection adopts the VisuShrink method, that is Where σ is the standard deviation of noise and N is the number of data points. Finally, the cleaned and denoised data is normalized using the minimum-maximum normalization method, i.e. x′=(xx min ) / (x max -x min ), where x is the original data, x min and x max are the minimum and maximum values ​​of the data set, respectively. The purpose of this step is to eliminate noise and outliers in the original data and unify the data to the same scale in preparation for subsequent feature extraction.

[0097] The specific implementation of step S50 is: first extract the overall features. Calculate the mean, variance, peak, skewness and kurtosis of the resistivity data of the entire anchor pile. The mean reflects the overall resistivity level, the variance indicates the discreteness of the data, the peak indicates the maximum resistivity, the skewness reflects the asymmetry of the data distribution, and the kurtosis indicates the sharpness of the data distribution. These features can reflect the resistivity distribution characteristics of the anchor pile as a whole. Then extract the periodic features. Use fast Fourier transform (FFT) to perform spectrum analysis on the resistivity data to identify the main periodic components. For each identified period, calculate its period length, fluctuation amplitude, phase difference and frequency. At the same time, use the sliding time window method to extract the time window features. The window length is set to three scales of 1%, 5% and 10% of the total data length, and the overlap rate of each window is 50%. Calculate the mean, standard deviation, maximum value, minimum value, kurtosis and skewness in each window to form a time window feature vector. These features can reflect the changing characteristics of the resistivity data on different time scales. Finally, extract local features. The local anomaly factor (LOF) algorithm is used to detect local anomalies, and the LOF threshold is set to 1.5. For each detected anomaly, its position, amplitude (deviation from the local mean), duration (number of consecutive anomalies), and shape (characterized by the second-order derivative) are recorded. These features can reflect the local abnormal conditions of the anchor pile, such as uneven grouting, cracks, etc. The purpose of this step is to extract features that can fully reflect the quality of the anchor pile from the preprocessed resistivity data and provide input for subsequent model analysis.

[0098] The specific implementation method of step S60 is: first, the extracted feature vector is input into the pre-trained non-destructive testing model. The model adopts a multi-branch convolutional neural network structure, including an overall feature detection branch, a periodic feature detection branch, a local feature detection branch, a superposition attention module and a fusion branch. The overall feature detection branch uses a three-layer fully connected network, with 64, 32, and 16 nodes in each layer, and the activation function is ReLU. The periodic feature detection branch uses a one-dimensional convolutional neural network, including 3 convolutional layers (the convolution kernel sizes are 3, 5, and 7, respectively, and 64 filters in each layer) and 2 maximum pooling layers. The local feature detection branch uses an attention-enhanced convolutional neural network, including 2 convolutional layers and 1 self-attention layer. The superposition attention module adopts a multi-head self-attention mechanism with 8 heads. The fusion branch uses a two-layer fully connected network with 32 and 16 nodes, respectively, and finally a softmax layer. The input of the model is all the features extracted in step S50, and the output is a 5-dimensional vector, which respectively represents the degree of opening of the enlarged head (0-100%), the quality of grouting (excellent, qualified, unqualified), the integrity of the anchor section (complete, slightly damaged, severely damaged), the state of the free section (normal, slightly corroded, severely corroded) and the degree of deformation of the steel bar (no deformation, slightly deformed, severely deformed). The model training uses the cross entropy loss function and the Adam optimizer, with a learning rate of 0.001, a batch size of 64, and a training round number of 100. During the training process, the early stopping method is used to prevent overfitting, and the training is stopped if the loss on the validation set does not decrease for 5 consecutive rounds. The purpose of this step is to use the deep learning model to comprehensively analyze the extracted features and output a comprehensive evaluation result of the anchor pile quality.

[0099] The specific implementation method of step 11 is: first, design and produce a batch of flexible photovoltaic support cement grouting anchor pile samples containing a single quality defect. The quality defects include incomplete opening of the enlarged head (the opening degrees are 25%, 50%, and 75%, respectively), insufficient grouting (grouting rates are 60%, 70%, and 80%, respectively), cracks in the anchoring section (crack widths are 0.1 mm, 0.3 mm, and 0.5 mm, respectively), corrosion in the free section (corrosion depths are 0.5 mm, 1 mm, and 1.5 mm, respectively), and deformation of the steel bars (deformation angles are 5 degrees, 10 degrees, and 15 degrees, respectively). Three samples of different degrees are made for each defect, and each degree of sample is repeated 3 times, making a total of 45 samples. Then, these samples are processed and measured according to the method of steps S10-S50 to obtain the key features of each sample. Features include overall features (mean, variance, peak, skewness, kurtosis), periodic features (time window features, period length, period fluctuation amplitude, period phase difference, period frequency) and local features (abnormal point location, abnormal point amplitude, abnormal point duration, abnormal point shape). Finally, these features are paired with the corresponding quality defect type and degree labels to form the first feature data set. The purpose of this step is to establish the correspondence between single defects and resistivity features to provide basic data for subsequent composite defect analysis.

[0100] The specific implementation method of step 12 is: first, design and produce a batch of flexible photovoltaic support cement grouting anchor pile samples containing at least two quality defects. The defect combination takes into account common situations in actual engineering, such as the expansion head is not fully opened and the grouting is insufficient, the anchor section cracks and the free section corrodes at the same time. Three different defect degree combinations are selected for each combination, and each combination is repeated 3 times. A total of 60 composite defect samples are made. Then, these samples are also processed and measured according to the method of steps S10-S50 to obtain the key features of each sample. The feature extraction method is the same as step 11, including overall features, periodic features and local features. Finally, these features are paired with corresponding multiple quality defect types and degree labels to form a second feature data set. During the data processing process, special attention is paid to recording the effects of different defect combinations on resistivity characteristics, such as some defect combinations may produce nonlinear superposition effects. The purpose of this step is to establish a corresponding relationship between composite defects and resistivity characteristics, and provide more complex and more practical training data for subsequent defect identification and quantitative evaluation.

[0101] The specific implementation method of step 13 is: first, all single quality defects (a total of 5 types) are combined in pairs to obtain 10 defect pairs. Then, for each defect pair, a flexible photovoltaic support cement grouting anchor pile sample containing the defect pair is designed and produced. Three different degree combinations are considered for each defect pair, and each combination is repeated 3 times, making a total of 90 samples. For example, for the defect pair of "enlarged head is not fully opened" and "inadequate grouting", three combinations of (25% opening degree, 70% grouting rate), (50% opening degree, 80% grouting rate), and (75% opening degree, 60% grouting rate) can be designed. Then, these samples are processed and measured according to the method of steps S10-S50 to obtain the key features of each sample. The feature extraction method is the same as the previous two steps, including overall features, periodic features, and local features. Finally, these features are paired with the corresponding defect pair type and degree labels to form a third feature data set. In the data processing process, special attention is paid to the interaction between the two defects, such as whether a defect will cover or enhance the characteristic performance of another defect. The purpose of this step is to deeply study the mutual influence between different defects and provide an important basis for the subsequent establishment of a more accurate defect identification model.

[0102] The specific implementation of step 14 is as follows: First, for each defect pair, extract the corresponding third feature and two related first features. Then, calculate the difference between the defect pair feature and the single defect feature. The specific method is to use two metrics, cosine similarity and Euclidean distance. For the feature vector F ij (defective pair), F i and F j (single defect), calculate cosine similarity: And the Euclidean distance: d = || F ij -(F i +F j )||2. Next, define the relative influence where d max is the maximum Euclidean distance among all defect pairs. This relative influence reflects the degree of influence of the combination of two defects on the characteristics. The larger the value, the more significant the combination effect. The relative influence is calculated for each feature (overall feature, periodic feature, local feature) to obtain a comprehensive impact assessment. Finally, the relative influence of all defect pairs is normalized so that their sum is 1. The purpose of this step is to quantify the influence of different defect combinations on the resistivity characteristics and provide a data basis for the subsequent construction of the weight matrix.

[0103] The specific implementation of step 15 is as follows: First, a 5×5 feature influence weight matrix W is constructed based on the relative influence calculated in step 14. Each element W of the matrix ijRepresents the relative impact of defect i and defect j when combined. The matrix is ​​symmetrical, and the diagonal elements are 0 (indicating a single defect). Then, the matrix is ​​normalized so that the sum of each row and column is 1, that is, ∑ j W ij =1 and ∑ i W ij =1. This normalization ensures that the overall influence of different defects is equal. Next, the main eigenvectors of the matrix are analyzed using the eigenvalue decomposition method, and the eigenvectors corresponding to the first three largest eigenvalues ​​are selected as the main influence modes. These eigenvectors reflect the main modes of defect combinations and can be used for subsequent feature importance analysis. Finally, this weight matrix is ​​applied to the superposition attention module. Specifically, when calculating the attention weight, a regularization term based on the weight matrix is ​​added: Among them, λ is an adjustable hyperparameter used to control the influence of the weight matrix. Initially, λ is set to 0.1 and can be fine-tuned during the model training process. The purpose of this step is to incorporate the influence of defect combinations into the deep learning model and improve the model's ability to identify complex defect situations.

[0104] The specific implementation of step 16 is: first, integrate the first feature, second feature and third feature data sets obtained in steps 11, 12 and 13. For each sample, construct a feature vector containing all extracted features. The feature vector includes overall features (5 dimensions), periodic features (6 dimensions for each time scale window, a total of 18 dimensions) and local features (4 dimensions), totaling 27 dimensions. Then, according to the weight matrix constructed in step 15, the feature vector is weighted. The weighting method is to divide the feature vector into 5 groups (corresponding to 5 types of defects), each group of features is dot-multiplied with the corresponding row of the weight matrix, and then the results are connected into a new feature vector. This weighting method emphasizes the impact of different defect combinations on the features. Next, construct a label vector. For a single defect sample, the label is a 5-dimensional one-hot vector; for a compound defect sample, the label is a 5-dimensional multi-label vector, each element of which represents the degree of the corresponding defect (0 represents no defect, 1, 2, and 3 represent slight, medium, and severe defects, respectively). Finally, the weighted feature vector and the corresponding label vector are paired to form the final training data set. In order to balance the dataset, the minority class samples are oversampled and the SMOTE (Synthetic Minority Oversampling Technique) algorithm is used to generate synthetic samples. The final training dataset contains original samples and synthetic samples, with a total number of about 500 samples. The purpose of this step is to build a comprehensive and balanced training dataset to provide high-quality input data for subsequent model training.

[0105] The training steps of the non-destructive testing model are specifically:

[0106] Step 21: Data preparation and preprocessing. First, the training dataset constructed in step 16 is randomly divided into a training set (70%), a validation set (15%), and a test set (15%). The feature data is standardized using the Z-score method, i.e., x′ = (x-μ) / σ, where μ and σ are the mean and standard deviation of each feature, respectively. This step ensures that features of different scales can be treated fairly in the model.

[0107] Step 22: For the superimposed attention module, consider the feature influence weight matrix, apply residual connection and layer normalization to enhance the model's expressiveness and training stability. The specific implementation is to add residual connection after attention calculation: H i =LayerNorm(X i +MultiHead(X i )), where X i is the input, MultiHead(·) is the multi-head attention function, and LayerNorm(·) is the layer normalization. In the multi-head attention calculation, the weight matrix in step 15 is introduced: MultiHead(X) = Concat(head1,…,head h )W O , where head i =Attention(XW i Q ,XW i K ,XW i V ,W), where W is the feature influence weight matrix. This design can introduce prior knowledge about defect interactions while maintaining the flexibility of the model.

[0108] Step 23: Model training. Use the Adam optimizer for parameter optimization, and the initial learning rate is set to 0.001. The loss function uses weighted cross entropy loss to handle multi-label classification problems: Where N is the number of samples, y ij is the true label, p i j is the predicted probability, w j is the class weight, which is used to balance the importance of different defect types. Batch gradient descent is used, and the batch size is set to 64. To prevent overfitting, a Dropout layer is added between the fully connected layers, and the dropout rate is 0.5. At the same time, L2 regularization is used, and the regularization coefficient is 0.0001.

[0109] Step 24: Model validation and tuning. After each training epoch, evaluate the model performance on the validation set. Performance metrics include accuracy, precision, recall, and F1 score. If the F1 score on the validation set does not improve for 5 consecutive epochs, reduce the learning rate (multiply by 0.1). If there is no improvement for 10 consecutive epochs, trigger the early stopping mechanism. Use the grid search method to tune the hyperparameters, including learning rate, batch size, dropout rate, etc. Use 5-fold cross validation to evaluate the performance of each set of hyperparameter configurations.

[0110] Step 25: Model testing and deployment. Retrain the model on the full training set using the best hyperparameter configuration. Then evaluate the final model performance on the test set. For each defect type, calculate the confusion matrix and ROC curve to analyze the strengths and weaknesses of the model. Finally, save the trained model parameters.

[0111] Specifically, the principle of the present invention is:

[0112] 1. Multi-point electrode layout: During the anchor pile construction process, multiple measuring electrodes are laid out at key locations such as the enlarged head, anchoring section, and free section. This distributed electrode layout method enables the capture of resistivity information at different locations, laying the foundation for subsequent feature extraction and model analysis. At the same time, by recording the specific location of each electrode, it also provides an important basis for the interpretation of the results.

[0113] 2. Resistivity data collection: After the anchor construction is completed, the resistivity of these pre-buried electrodes is measured using professional electrical instruments, and relevant parameters such as electrode position, power supply current, measurement voltage, ambient temperature and soil moisture content are recorded. These precise on-site measurement data provide a reliable basis for subsequent data preprocessing and feature extraction.

[0114] 3. Multi-scale feature extraction: Extract multi-scale features including overall features, periodic features and local features from the pre-processed resistivity data. Overall features reflect the global statistical characteristics of resistivity data, such as mean, variance, peak value, etc.; periodic features capture the periodic patterns in resistivity data, such as time window features, period length and phase difference, etc.; local features describe the abnormal local features in resistivity data, such as abnormal point location, amplitude and duration, etc. This multi-angle feature extraction makes it possible to comprehensively characterize the quality status of the anchor pile.

[0115] 4. Multi-branch neural network model: The present invention adopts a multi-branch convolutional neural network structure, which includes an overall feature detection branch, a periodic feature detection branch and a local feature detection branch. The three branch networks extract feature representations of different scales respectively, and then perform weighted fusion through a superposition attention module, and finally output the anchor pile quality vector including multiple dimensions such as the enlarged head, grouting, anchoring section, free section and steel bar. This multi-branch structure can make full use of the complementary information of different features and improve the accuracy and robustness of non-destructive testing.

[0116] 5. Feature influence weight matrix: During the model training process, the present invention introduces a feature influence weight matrix to characterize the relative influence of different quality defects on the features. This weight matrix can not only guide the importance of different features in the model, but also be used to initialize and constrain the attention weights of the superposition attention module, so that the model can better capture the key features corresponding to different defects. This modeling method based on domain knowledge greatly improves the generalization ability of the model.

[0117] In summary, the nondestructive testing method proposed in the present invention makes full use of the rich information contained in the resistivity data, and through key technical means such as multi-point electrode layout, multi-scale feature extraction and multi-branch neural network modeling, a comprehensive assessment of the quality status of cement grouting anchor piles of photovoltaic brackets is achieved.

[0118] In order to better understand and implement the present invention, an embodiment of a specific application scenario of the present invention is provided below: A photovoltaic power station adopts a flexible photovoltaic support structure, and its foundation is fixed on the foundation in the form of cement grouting anchor piles. In order to ensure the long-term safety and reliability of the photovoltaic support structure, the power station operator decides to use the non-destructive testing method based on resistivity measurement proposed in the present invention to conduct a comprehensive assessment of the cement grouting anchor piles.

[0119] First, during the anchor pile construction process, 10 measuring electrodes were arranged at key locations such as the enlarged head, anchoring section, and free section. The specific location information of each electrode is shown in Table 1. This distributed electrode arrangement method enables the resistivity information of different locations to be accurately captured.

[0120] Table 1 Anchor pile electrode layout information

[0121] Electrode number X coordinate (m) Y coordinate (m) Z coordinate (m) Location 1 5.2 3.1 0.5 Enlarged head 2 5.3 3.2 0.6 Enlarged head … … … … … 10 5.9 3.7 1.2 Free segment

[0122] After the anchor pile construction was completed, the resistivity of these pre-buried electrodes was measured using professional electrical instruments. During the one-week measurement process, the power supply current, measurement voltage, ambient temperature, soil moisture content and other parameters of each electrode were recorded, and the corresponding resistivity values ​​were calculated. Table 2 shows some of the measurement data.

[0123] Table 2 Anchor pile resistivity measurement data

[0124]

[0125]

[0126] Based on these raw measurement data, they were first preprocessed, including data cleaning, normalization, and noise reduction. Specifically, obviously abnormal data points, such as extreme values ​​caused by instrument failure or environmental interference, were removed. Then, data of different dimensions (such as resistivity, temperature, water content, etc.) were mapped to the [0,1] interval to avoid certain features dominating in subsequent model training. Finally, a moving average filter was applied to reduce the noise of the resistivity data to improve the signal-to-noise ratio.

[0127] After completing data preprocessing, multi-scale features including overall features, periodic features and local features were extracted from the resistivity data.

[0128] Overall features include:

[0129] 1. Mean resistivity Where N is the number of measurement points, ρ i is the resistivity value of the i-th measurement point.

[0130] 2. Resistivity variance

[0131] 3. Resistivity peak

[0132] 4. Resistivity skewness

[0133] 5. Resistivity peak

[0134] Cycle characteristics include:

[0135] 1. Time window feature W j =[μ j ,σ j ,max j ,min j ,K j ,S k,j ], where μ j ,σ j ,max j ,min j ,K j ,S k,j are the mean, standard deviation, maximum, minimum, kurtosis and skewness of the resistivity in the jth time window, respectively.

[0136] 2. Cycle length Where M is the number of cycles identified, t j is the start time of the jth cycle.

[0137] 3. Cycle fluctuation range where ρ max,j and ρ min,j are the maximum and minimum resistivity of the jth period, respectively.

[0138] 4. Cycle phase difference where t peak is the peak time, t ref For reference time.

[0139] 5. Cycle frequency

[0140] Local features include:

[0141] 1. Location of abnormal points Where θ is the abnormal threshold and the default value is 3.

[0142] 2. Abnormal point amplitude

[0143] 3. Duration of abnormal point

[0144] 4. Outlier shape factor in is the second-order derivative of resistivity.

[0145] These extracted multi-scale features constitute the input data for this nondestructive testing.

[0146] In order to evaluate the relative impact of these features on different quality defects, several groups of cement grouting anchor pile samples with single defects and multiple defects were specially manufactured. By testing these samples and extracting features, the feature influence weight matrix W was calculated. ij , the formula is:

[0147]

[0148] Among them, W ij is the relative impact weight of defect i and defect j; F ij is the third feature including defects i and j; F i and F j are the first features containing only defects i and j respectively; ||·||2 represents the L2 norm.

[0149] Figure 2The calculated feature influence weight matrix is ​​given, and it can be seen that the degree of influence of different defects on the features varies greatly. For example, the incomplete opening of the enlarged head has the greatest impact on the overall features, while the free segment corrosion mainly affects the local features. These results provide a basis for the initialization and constraints of the attention mechanism in model design.

[0150] Based on the extracted multi-scale features and feature influence weight matrix, a multi-branch convolutional neural network structure is used as a non-destructive testing model. The model includes an overall feature detection branch, a periodic feature detection branch and a local feature detection branch. The model is fused through a superposition attention module to finally output the anchor pile quality vector with multiple dimensions, including the opening degree of the enlarged head, the grouting quality, the integrity of the anchor section, the state of the free section and the deformation of the steel bar. The overall feature detection branch adopts a multi-layer fully connected network structure to extract the global resistivity features; the periodic feature detection branch adopts a one-dimensional convolutional neural network to identify the periodic patterns in the resistivity data; the local feature detection branch adopts an attention-enhanced convolutional network to effectively capture local abnormal features. The outputs of these three branches are weighted fused through the superposition attention module and finally input into the fully connected network and the softmax layer to give the evaluation results of each quality indicator. In the model training process, the attention weights of the superposition attention module are first initialized and constrained according to the feature influence weight matrix, so that the model can better capture the key features corresponding to different defects. Then, the model parameters are optimized by the mini-batch stochastic gradient descent method, and the loss function adopts the cross entropy loss. To prevent overfitting, early stopping and L2 regularization techniques were applied. The model performance was monitored on the validation set until the model converged or the preset number of training rounds was reached, and the optimal model parameters were finally saved. By implementing this non-destructive testing method, operators can not only comprehensively evaluate the quality status of cement grouting anchor piles of existing photovoltaic brackets, timely discover and diagnose potential defects, but also provide a reliable quality control reference for future new projects. Compared with traditional destructive testing methods, this method has significant advantages: no destructive operation is required on the anchor piles, the testing cycle is shorter, the cost is lower, and it can provide more comprehensive and accurate quality assessment results. This not only helps to improve the safety and service life of photovoltaic bracket structures, but also provides important guarantees for the long-term healthy operation of photovoltaic power stations.

[0151] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A nondestructive testing method for cement grouting anchor piles of flexible photovoltaic support, characterized in that: The following steps are involved: S10, preparation stage: insulation and anti-corrosion treatment of the steel bars and enlarged heads of the anchor piles to ensure the reliability of electrode layout; S20, electrode layout: during the anchor bolt construction process, multiple electrodes are laid out at the enlarged head, anchoring section and free section, and positioning electrode layout is performed; S30, resistivity measurement: after the anchor construction is completed, use the electrical instrument to measure the resistivity and record the measurement data; S40, data preprocessing: cleaning, normalizing and noise reduction of the measured resistivity data; S50, feature extraction: extracting key features from the processed resistivity data, including overall features, periodic features and local features; S60, model analysis: input the key features into a pre-trained non-destructive testing model, and output the quality vector of the anchor pile.

2. According to claim 1, a nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support is characterized in that: The measurement data include electrode position, power supply current, measurement voltage, calculated resistivity, measurement time, ambient temperature, and soil moisture content.

3. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 2, characterized in that: The overall characteristics include mean, variance, peak, skewness, and kurtosis.

4. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 3, characterized in that: The periodic characteristics are period-related characteristics of each type of continuous resistivity data of different time lengths, including time window characteristics, period length, period fluctuation amplitude, period phase difference, and period frequency of the continuous resistivity data of different time lengths.

5. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 4, characterized in that: The time window characteristics of the continuous resistivity data of different time lengths specifically refer to segmenting the resistivity data using a sliding time window and extracting statistical characteristics within each time window, including mean, standard deviation, maximum value, minimum value, kurtosis and skewness.

6. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 5, characterized in that: The local features are local abnormal features in the resistivity data, including abnormal point positions, abnormal point amplitudes, abnormal point durations, and abnormal point shapes.

7. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 6, characterized in that: The mass vector is specifically a multidimensional vector including the opening degree of the enlarged head, the grouting quality, the integrity of the anchoring section, the state of the free section and the deformation of the steel bar.

8. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 7, characterized in that: The nondestructive testing model adopts a multi-branch convolutional neural network structure, including an overall feature detection branch, a periodic feature detection branch, a local feature detection branch, a superimposed attention module and a fusion branch.

9. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 8, characterized in that: The overall feature detection branch is used to extract global resistivity features, the input is an overall feature vector, the output is a global feature representation, and the structure is a multi-layer fully connected network; The periodic feature detection branch is used to identify the periodic pattern of resistivity data, the input is a periodic feature sequence, the output is a periodic feature representation, and the structure is a one-dimensional convolutional neural network; The local feature detection branch is used to capture local abnormal features, the input is a local feature sequence, the output is a local feature representation, and the structure is an attention-enhanced convolutional neural network; The superposition attention module is used to perform weighted fusion on the outputs of different feature branches, and its structure is a multi-head self-attention mechanism; The fusion branch is used to integrate all feature representations and output the final result. The input is the feature representation after weighted fusion, and the output is the anchor pile mass vector. The structure is a fully connected network and a softmax layer.

10. A nondestructive testing method for cement grouting anchor piles of a flexible photovoltaic support according to claim 9, characterized in that: The steps of establishing the training data set of the non-destructive testing model are specifically: Step 11, obtaining multiple groups of cement grouting anchor piles of flexible photovoltaic brackets with a single quality defect, and using steps S10-S50 to obtain the key features of each group of anchor piles, recorded as the first feature; the quality defects specifically include incomplete opening of the enlarged head, insufficient grouting, cracks in the anchoring section, corrosion in the free section, and deformation of the steel bar; Step 12, obtaining multiple groups of cement grouting anchor piles of flexible photovoltaic supports with at least two quality defects, and using steps S10-S50 to obtain key features of each group of anchor piles, recorded as the second feature; Step 13, group all types of quality defects into pairs to form defect pairs, obtain each type of flexible photovoltaic support cement grouting anchor pile with defect pairs, and use steps S10-S50 to obtain the key features of each group of anchor piles, which are recorded as the third features; Step 14: According to each defect pair, the relative influence of the two defects in the defect pair on the feature is calculated according to the corresponding third feature and the two first features; Step 15: According to the relative influence of all defect pairs, a feature influence weight matrix is ​​established to adjust the importance of different features in the model and to initialize and constrain the attention weight of the superposition attention module; Step 16: Establish a training data set based on the weight matrix, the first feature, the second feature, and the third feature, wherein the input of the training is the key feature, and the output of the training is the corresponding quality defect label.

Images