Damage identification method of concrete square piles based on adaptive multi-channel convolutional neural network

Abstract

The present invention discloses a method for identifying damage of concrete square piles based on an adaptive multi-channel convolutional neural network. The method comprises the following steps: step 1, using a sensor array arranged on the concrete square pile to collect signals obtained by striking with an exciting hammer, and making corresponding category labels; step 2, building an AMCCNN model, and using the labeled data collected in step 1 to train the model until the model converges; step 3, applying the trained AMCCNN model to practice, performing state diagnosis on actual concrete square piles, timely identifying square piles in abnormal states, and reducing the occurrence of accidents. The method of the present invention can quickly and accurately identify square piles in abnormal states, providing a favorable basis for taking effective measures before an accident occurs, and has good practical application value.

Description

Technical Field

[0001] The invention relates to the field of concrete square pile damage state recognition and evaluation, and in particular to a concrete square pile damage recognition method based on an adaptive multi-channel convolutional neural network. Background Art

[0002] Concrete foundation piles are a type of columnar component that transfers the load of the upper structure to the soil layers of the foundation along the depth through the friction of the pile body, and to the rock layer with higher strength and lower compressibility at the bottom through the pile bottom end bearing. Its function is to reduce the adverse effects of shallow soft strata on the upper structure and provide reliable bearing capacity and deformation control capabilities for the upper structure. With the development of infrastructure construction, complex geological conditions are encountered more and more frequently during engineering construction, and people have put forward higher requirements for the functions of pile foundations.

[0003] Among many detection methods, the low-strain reflection wave method has become the most effective pile foundation detection method in the engineering community at home and abroad due to its advantages of rapidity, accuracy, and convenience. However, it requires a lot of assumptions, and is subject to the superposition of test noise, pile structure parameter variability, excitation parameter variability, and geological parameter variability in actual engineering. The defect analysis results of the low-strain reflection wave method are uncertain, which may cause misjudgment of pile integrity classification.

[0004] The current neural network has been widely used in various fields due to its excellent adaptive nonlinear expression ability. In the field of concrete pile foundation, some people have used neural networks to identify multiple parameters and nonlinear optimization of pile foundation dynamic measurement and predict the relationship curve between single pile load and settlement. Some people have also used neural networks to identify the damage state of pile foundation. All of the above methods have achieved good engineering application results.

[0005] Through research on relevant literature and analysis of the current research status, it is found that there are currently two major problems in the field of pile foundation status recognition using neural network models: 1. Because the collected model training data is affected by factors such as actual noise, variability of pile structure parameters and variability of excitation parameters, the characteristics of the original fault information in the pile foundation are not obvious, resulting in low recognition accuracy of the model after training; 2. The characteristics of the collected pile foundation data are multi-layered and complex, and the current research model is insufficient in multi-feature fusion and expression, which will directly determine the generalization ability of the model application. Summary of the invention

[0006] To solve the above problems, the present invention proposes a concrete square pile damage identification method based on an adaptive multi-channel convolutional neural network based on a discrete wavelet threshold denoising algorithm and a convolutional neural network (CNN). For problem 1, this patent proposes an adaptive discrete wavelet threshold denoising algorithm, which overcomes the problems of denoising glitches and incomplete denoising caused by using a global threshold for denoising in the traditional wavelet threshold denoising algorithm. The proposed convergence criterion function better discretizes and aggregates the wavelet coefficient group, and finally adaptively determines the denoising threshold of each layer, expanding from the local optimum to the global optimum, thereby enhancing the signal characteristics. For problem 2, this patent designs a multi-channel convolutional neural network, which can gradually fuse multi-channel and multi-level features. Through the designed multi-channel fusion pooling layer, different convolutional layers are fused. The proposed fusion pooling algorithm can map the convolutional layer features to the mean-maximum range, so that the network can learn dynamic feature data, thereby improving the learning ability of the network and the generalization performance of the model application. To achieve this purpose, the present invention provides a concrete square pile damage identification method based on an adaptive multi-channel convolutional neural network, the specific steps are as follows, and the characteristics are as follows:

[0007] Step 1: AMCCNN training data collection. The sensor array arranged on the concrete square piles collects the signals obtained by the vibration hammer. Six types of square pile integrity states are collected, namely: fracture, enlarged diameter, segregation, reduced diameter, multiple defects and intact, and corresponding category labels are made;

[0008] Step 2: Offline training of the AMCCNN model: Build the AMCCNN model and use the labeled data collected in step 1 to train the model until the model converges;

[0009] Step 3, online application of the training model: Apply the trained AMCCNN model to practice, diagnose the status of actual concrete piles, identify piles in abnormal status in time, and reduce the occurrence of accidents.

[0010] 2. According to the method for concrete square pile damage identification based on adaptive multi-channel convolutional neural network in claim 1, the algorithm model AMCCNN is characterized in that: the sensor array arrangement rule for collecting the square pile state data in step 1 is: four sensors are arranged on the square pile plane, the sensors are distributed in a square shape and the center of the sensor array coincides with the center of the square pile, and the distance between each sensor is the length of the corresponding square pile side.

[0011] 3. According to the method for concrete square pile damage identification based on adaptive multi-channel convolutional neural network in claim 1, the algorithm model AMCCNN is characterized in that: the specific steps of offline training of the AMCCNN model in step 2 are:

[0012] Step 2.1. An adaptive discrete wavelet threshold denoising algorithm is proposed to filter and enhance the features of the signals collected by each sensor, and an adaptive threshold is selected for each layer in the wavelet decomposition. The selection criteria are as follows:

[0013] The coefficient sequence P of the kth layer obtained after wavelet decomposition k For example, under the prior threshold λ k Under the condition of k Divided into two groups P pre,k and P nxt,k ,Right now:

[0014] P pre,k ={x k,min ,x k,min +1,...,λ k -1}

[0015] P nxt,k ={λ k ,λ k +1,...,x k,max -1,x k,max}

[0016] x k,min and x k,max They represent the minimum and maximum values ​​of the wavelet coefficients of the kth layer respectively.

[0017] In the iterative calculation, let λ k The initial value of is:

[0018]

[0019] u pre,k and u nxt,k Respectively represent P pre,k and P nxt,k The mean value, s pre,k and nxt,k Respectively represent P pre,k and P nxt,k The standard deviation of P pre,k and P nxt,k , while making P pre,k and P nxt,k Each of them is cohesive. This patent defines the convergence criterion function. The k-th convergence criterion function J k The expression is as follows:

[0020]

[0021] Continue to update λ in the next iteration k , that is, k =(upre,k +u nxt,k ) / 2, and then re-divide the coefficient sequence according to the above principle, and then solve the mean and standard deviation until the convergence criterion function value of this iteration is less than the previous one, that is, J n >J n+1 .

[0022] Step 2.2. Use the FFT algorithm to solve the frequency domain curve of the filtered and enhanced signal obtained in step 2.1., where the sampling rate is set to 12800 Hz and the sampling point is set to 4960;

[0023] Step 2.3. Build a multi-channel convolutional neural network, where the network architecture is: multi-channel convolution layer 1-multi-channel fusion pooling layer 1-multi-channel convolution layer 2-fusion pooling layer 2-Flatten layer-fully connected layer 1-fully connected layer 2-SOFTMAX layer. Among them, multi-channel convolution layer 1 (channels 1, 2, 3 and 4) processes the data of step 2.2; multi-channel fusion pooling layer 1 (channels 1 and 2) fuses the data of channels 1 and 3, channels 2 and 4 of multi-channel convolution layer 1 respectively; channels 1 and 2 of multi-channel convolution layer 2 process the data of channels 1 and 2 of multi-channel fusion pooling layer 1 respectively; fusion pooling layer 2 fuses the features of channels 1 and 2 of multi-channel convolution layer 2, and then flattens the features in a Flattern manner, and then performs feature dimensionality reduction processing through fully connected layers 1 and 2, and finally uses the SOFTMAX layer to realize the identification of pile foundation damage status. Among them, the processing details of fusion pooling (FusionPooling) of the fusion pooling layer are:

[0024] For the convolutional layer feature c with specification s×t ij (1≤i≤s,1≤j≤t), calculate the corresponding mean c ave and the maximum value c max ,Right now:

[0025]

[0026] c max =max(c ij )(1≤i≤s,1≤j≤t)

[0027] Then calculate the corresponding fusion pooling result c fuP , the calculation expression is:

[0028]

[0029] Step 2.4. Continuously iterate the network training until the network model converges. The loss function of network training is cross entropy loss, the network parameter optimization algorithm is Adam, and the condition for iterative termination is that the loss function is less than 1e-3.

[0030] The present invention is based on an adaptive multi-channel convolutional neural network concrete square pile damage identification method. Based on the low-strain signals of pile foundations with different integrity collected by the sensor array, the collected signals are first adaptively filtered and feature enhanced by proposing an adaptive discrete wavelet threshold denoising algorithm, and then a multi-channel convolutional neural network is proposed to achieve data dimension reduction and feature extraction through feature fusion of different channels, and finally accurately identify the integrity status of the pile foundation. The overall beneficial effects are:

[0031] 1. The adaptive discrete wavelet threshold denoising algorithm proposed in this patent overcomes the problems of denoising glitches and incomplete denoising caused by using global threshold for denoising in the traditional wavelet threshold denoising algorithm. The proposed convergence criterion function better discretizes and aggregates the wavelet coefficient group, and finally adaptively determines the denoising threshold of each layer, expanding from the local optimum to the global optimum, thereby enhancing the signal characteristics.

[0032] 2. Through the designed multi-channel convolutional neural network, multi-channel and multi-level features can be gradually integrated. Through the designed multi-channel fusion pooling layer, features of different convolutional layers can be fused. The proposed fusion pooling algorithm can map the convolutional layer features to the mean-maximum range, so that the network can learn dynamic feature data, thereby improving the network's learning ability and the generalization performance of the model application. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the present invention;

[0034] Figure 2 is a schematic diagram of the diagnostic process in this patent;

[0035] Figure 3 This is the network architecture diagram of the AMCCNN model in this patent. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0037] The present invention proposes a concrete square pile damage identification method based on an adaptive multi-channel convolutional neural network, which aims to quickly and accurately identify square piles in abnormal states, discover hidden dangers in advance, reduce the occurrence of accidents, and reduce casualties. Figure 1 It is a flow chart of the present invention. It can be seen from the figure that the specific implementation steps of the method are:

[0038] Step 1: AMCCNN training data collection. The sensor array arranged on the concrete square piles collects the signals obtained by the vibration hammer. Six types of square pile integrity states are collected, namely: fracture, enlarged diameter, segregation, reduced diameter, multiple defects and intact, and corresponding category labels are made;

[0039] Step 2: Offline training of the AMCCNN model: Build the AMCCNN model and use the labeled data collected in step 1 to train the model until the model converges;

[0040] Step 3, online application of the training model: Apply the trained AMCCNN model to practice, diagnose the status of actual concrete piles, identify piles in abnormal status in time, and reduce the occurrence of accidents.

[0041] The specific steps of offline training of the AMCCNN model in step 2 are:

[0042] Step 2.1. An adaptive discrete wavelet threshold denoising algorithm is proposed to filter and enhance the features of the signals collected by each sensor, and an adaptive threshold is selected for each layer in the wavelet decomposition. The selection criteria are as follows:

[0043] The coefficient sequence P of the kth layer obtained after wavelet decomposition k For example, under the prior threshold λ k Under the condition of k Divided into two groups P pre,k and P nxt,k ,Right now:

[0044] P pre,k ={x k,min ,x k,min +1,...,λ k -1}

[0045] P nxt,k ={λ k ,λ k +1,...,x k,max -1,x k,max}

[0046] x k,min and x k,max They represent the minimum and maximum values ​​of the wavelet coefficients of the kth layer respectively.

[0047] In the iterative calculation, let λ k The initial value of is:

[0048]

[0049] u pre,k and u nxt,k Respectively represent P pre,k and P nxt,k The mean value, s pre,k and nxt,k Respectively represent P pre,k and P nxt,k The standard deviation of P pre,k and Pnxt,k , while making P pre,k and P nxt,k Each of them is cohesive. This patent defines the convergence criterion function. The k-th convergence criterion function J k The expression is as follows:

[0050]

[0051] Continue to update λ in the next iteration k , that is, k =(u pre,k +u nxt,k ) / 2, and then re-divide the coefficient sequence according to the above principle, and then solve the mean and standard deviation until the convergence criterion function value of this iteration is less than the previous one, that is, J n >J n+1 .

[0052] Step 2.2. Use the FFT algorithm to solve the frequency domain curve of the filtered and enhanced signal obtained in step 2.1., where the sampling rate is set to 12800 Hz and the sampling point is set to 4960;

[0053] Step 2.3. Build a multi-channel convolutional neural network. The processing details of the fusion pooling of the fusion pooling layer are as follows:

[0054] For the convolutional layer feature c with specification s×t ij (1≤i≤s,1≤j≤t), calculate the corresponding mean c ave and the maximum value c max ,Right now:

[0055]

[0056] c max =max(c ij )(1≤i≤s,1≤j≤t)

[0057] Then calculate the corresponding fusion pooling result c fuP , the calculation expression is:

[0058]

[0059] Step 2.4. Continuously iterate the network training until the network model converges. The loss function of network training is cross entropy loss, the network parameter optimization algorithm is Adam, and the condition for iterative termination is that the loss function is less than 1e-3.

[0060] Figure 2This is a schematic diagram of the diagnostic process in this patent. As can be seen from the figure, the pile foundation is struck by a vibration hammer, and then the sensor array and low-strain detector are used to collect data, and then the integrity status of the pile foundation is identified using the trained AMCCNN model. The sensor array is arranged on the square pile plane, the sensors are distributed in a square shape, and the center of the sensor array coincides with the center of the square pile. The distance between each sensor is the length of the corresponding square pile side.

[0061] Figure 3 This is the network architecture diagram of the AMCCNN model in this patent. It can be seen from the figure that the specific network architecture is: multi-channel convolution layer 1-multi-channel fusion pooling layer 1-multi-channel convolution layer 2-fusion pooling layer 2-Flatten layer-fully connected layer 1-fully connected layer 2-SOFTMAX layer. The data collected by the sensor array is input into the multi-channel convolution layer 1 (channels 1, 2, 3 and 4); the multi-channel fusion pooling layer 1 (channels 1 and 2) fuses the data of channels 1 and 3, channels 2 and 4 of the multi-channel convolution layer 1 respectively; channels 1 and 2 of the multi-channel convolution layer 2 process the data of channels 1 and 2 of the multi-channel fusion pooling layer 1 respectively; the fusion pooling layer 2 fuses the features of channels 1 and 2 of the multi-channel convolution layer 2, and then flattens the features in a Flattern manner, and then performs feature dimensionality reduction processing through fully connected layers 1 and 2, and finally uses the SOFTMAX layer to realize the identification of the damage state of the pile foundation. The AMCCNN model can better fuse multi-channel and multi-level features through the designed multi-channel fusion pooling layer, thereby improving the learning ability of the network and the generalization performance of the model application.

Claims

1. A method for identifying damage of concrete square piles based on an adaptive multi-channel convolutional neural network, characterized by: Step 1: AMCCNN training data collection. The sensor array arranged on the concrete square piles collects the signals obtained by the vibration hammer. Six types of square pile integrity states are collected, namely: fracture, enlarged diameter, segregation, reduced diameter, multiple defects and intact, and corresponding category labels are made; The sensor array arrangement rule for collecting the state data of the square pile in step 1 is as follows: four sensors are arranged on the plane of the square pile, the sensors are distributed in a square shape and the center of the sensor array coincides with the center of the square pile, and the distance between each sensor is the length of the corresponding square pile side. Step 2: Offline training of the AMCCNN model: Build the AMCCNN model and use the labeled data collected in step 1 to train the model until the model converges; The specific steps for offline training of the AMCCNN model in step 2 are: Step 2.

1. An adaptive discrete wavelet threshold denoising algorithm is proposed to filter and enhance the features of the signals collected by each sensor, and an adaptive threshold is selected for each layer in the wavelet decomposition. The selection criteria are as follows: The coefficient sequence P of the kth layer obtained after wavelet decomposition k For example, under the prior threshold λ k Under the condition of k Divided into two groups P pre,k and P nxt,k ,Right now: P pre,k ={x k,min ,x k,min +1,...,l k -1} P nxt,k ={λ k ,l k +1,...,x k,max -1.x k,max } x k,min and x k,max Respectively represent the minimum and maximum values ​​of the k-th layer wavelet coefficients; In the iterative calculation, let λ k The initial value of is: u pre,k and u nxt,k Respectively represent P pre,k and P nxt,k The mean value, s pre,k and nxt,k Respectively represent P pre,k and P nxt,k The standard deviation of P pre,k and P nxt,k , while making P pre,k and P nxt,k Each of them is cohesive, and the convergence criterion function is defined. The k-th convergence criterion function J k The expression is as follows: Continue to update λ in the next iteration k , that is, k =(u pre,k +u nxt,k ) / 2, and then re-divide the coefficient sequence according to the above principle, and then solve the mean and standard deviation until the convergence criterion function value of this iteration is less than the previous one, that is, J n >J n+1 ; Step 2.

2. Use the FFT algorithm to solve the frequency domain curve of the filtered and enhanced signal obtained in step 2.1., where the sampling rate is set to 12800 Hz and the sampling point is set to 4960; Step 2.

3. Build a multi-channel convolutional neural network, where the network architecture is: multi-channel convolution layer 1-multi-channel fusion pooling layer 1-multi-channel convolution layer 2-fusion pooling layer 2-Flatten layer-fully connected layer 1-fully connected layer 2-SOFTMAX layer, where multi-channel convolution layer 1, namely channels 1, 2, 3 and 4, processes the data of step 2.2; multi-channel fusion pooling layer 1, namely channels 1 and 2, fuses the data of channels 1 and 3, channels 2 and 4 of multi-channel convolution layer 1 respectively; channels 1 and 2 of multi-channel convolution layer 2 process the data of channels 1 and 2 of multi-channel fusion pooling layer 1 respectively; fusion pooling layer 2 fuses the features of channels 1 and 2 of multi-channel convolution layer 2, and then flattens the features in a Flatten manner, and then performs feature dimensionality reduction processing through fully connected layers 1 and 2, and finally uses the SOFTMAX layer to realize the recognition of pile foundation damage status, where the processing details of the fusion pooling of the fusion pooling layer are as follows: For the convolutional layer feature c with specification s×t ij (1≤i≤s,1≤j≤t), calculate the corresponding mean c ave and the maximum value c max ,Right now: c max =max(c ij )(1≤i≤s,1≤j≤t) Then calculate the corresponding fusion pooling result c fuP , the calculation expression is: Step 2.

4. Continuously iterate the network training until the network model converges. The loss function of network training is cross entropy loss, the network parameter optimization algorithm is Adam, and the condition for iterative termination is that the loss function is less than 1e-3; Step 3, online application of the training model: Apply the trained AMCCNN model to practice, diagnose the status of actual concrete piles, identify piles in abnormal status in time, and reduce the occurrence of accidents.

Images