Method for detecting durability of maritime work concrete under multi-field coupling condition by adopting machine vision

Abstract

The invention discloses a method for detecting the durability of maritime work concrete under a multi-field coupling condition by adopting machine vision. The detection method provided by the invention comprises an ocean immersion experiment and a wind tunnel experiment, respectively simulating a corrosion degradation process in an ocean environment and mutually verifying. The improved Attention U-Net deep full convolutional network is adopted, an attention gate module is further simplified in the attention U-Net, gating support is selected from output of a reverse direction layer, the simplification not only retains effectiveness of an attention mechanism, but also reduces calculation complexity, and the model can be subjected to high-precision training through a small amount of data.

Description

Technical Field

[0001] The invention relates to the field of machine vision and structural safety and durability of marine engineering concrete, and in particular to a method for detecting the durability of marine engineering concrete under multi-field coupling conditions by using machine vision. Background Art

[0002] With the development of global marine resources and coastal engineering, marine concrete has become a key material for building marine structures. It is widely used in infrastructure construction such as offshore wind power foundations, port terminals, and offshore platforms. However, the marine environment is complex and harsh. The long-term physical degradation of wave scouring, dry-wet cycles, freeze-thaw cycles, and ultraviolet radiation, as well as the corrosive ions, salt spray, carbonization reactions in seawater, and chemical reactions such as alkali-aggregate reactions inside concrete have jointly led to a gradual decline in the performance of concrete materials.

[0003] Concrete surface damage detection based on computer vision is a hot research area in computer science and civil engineering. For example, edge detection, morphological operations, thresholds, and Gabor filtering have been widely used to identify and quantify concrete surface cracks. With the development of deep learning, models such as convolutional neural networks (CNNs) and artificial neural networks (BNs) deep residual networks (ResNets) have been successfully applied to concrete crack detection. These machine learning-based methods can not only improve the accuracy of crack detection, but also classify crack morphology. Moreover, with the increase in data volume, the accuracy and robustness of these models are further enhanced.

[0004] There are already a wealth of technical methods for concrete flaw detection, including electrical quantity testing theory that studies the damage of concrete by monitoring its resistance, and acoustic quantity testing technology that detects micro-damage in concrete structures by spectrum and time nonlinear methods. In terms of concrete flaw detection technology, the damage model is mainly aimed at damage under mechanical loads, and is not directly related to corrosion and degradation damage in marine environments. There is also a lot of work accumulation on how temperature, salt, humidity and power affect the durability of marine engineering concrete. Chloride ions (Cl-) diffuse into the structure through the capillary pores of concrete, which will cause steel corrosion and lead to a decrease in structural strength. The corrosion ability of chloride ions is closely related to concentration. The higher the chloride ion concentration, the faster the corrosion rate. High temperature will accelerate the evaporation of water inside the concrete, resulting in an increase in structural cracks and a decrease in strength. At the same time, high temperature will also accelerate the diffusion and penetration rate of chloride ions, further aggravating steel corrosion. Chloride ions will accelerate the penetration of concrete structures under repeated dynamic loads. As the dynamic load changes, the internal microcracks of concrete gradually increase, providing more channels for chloride ion diffusion and accelerating steel corrosion. In a hot, humid and high-salt environment, the peak chloride ion concentration in concrete appears near the surface, and corrosion products tend to concentrate on the surface of the steel bars, causing a sharp drop in structural strength.

[0005] Traditional environmental factor coupling and machine vision methods cannot meet the durability assessment of marine concrete in harsh marine service environments:

[0006] 1) In terms of concrete flaw detection technology, the damage model is mainly aimed at the damage under mechanical load, which is not directly related to the corrosion degradation damage in the marine environment.

[0007] 2) Most of the existing durability tests on marine concrete are aimed at single or double-factor coupling experiments and numerical simulation analyses, while there is little research on the influencing mechanism and laws of the coupling of four factors: temperature, salt, humidity and power.

[0008] 3) Since it takes a long time for marine concrete to corrode and deteriorate, samples for machine vision learning are scarce, and the recognition accuracy is low in complex dynamic environments, especially in multi-field coupling environments. Summary of the invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for detecting the durability of marine concrete under multi-field coupling conditions using machine vision. In this patent, a multi-field coupling method of temperature, salinity, humidity and power is established, and corresponding multi-gradient change tests such as temperature, salinity, humidity, and wave making are set in the coupled wind farm equipment. By using an improved AttentionU-Net deep full convolutional network, the attention gate module is further simplified in the AttentionU-Net, and the gate support is selected from the output of the reverse direction layer. This simplification not only retains the effectiveness of the attention mechanism, but also reduces the computational complexity. The model can be trained with high precision using a small amount of data.

[0010] To achieve the above purpose, the technical solution designed by the present invention is as follows:

[0011] The present invention provides a method for detecting the durability of marine engineering concrete under multi-field coupling conditions using machine vision, including an ocean immersion test and a wind tunnel test, which respectively simulate the corrosion degradation process in a marine environment and verify each other.

[0012] Furthermore, the ocean immersion experiment comprises the following steps:

[0013] (1) Sample placement: The specimens are placed in a specific area on the seashore, which is divided into a full immersion area, a splash zone, and an atmospheric zone;

[0014] (2) Quantity configuration: 5 samples are placed in each area, for a total of 15;

[0015] (3) Regular sampling: 4 times a year, 3 samples each time, a total of 12 samples, 3 samples reserved as spare samples;

[0016] (4) Parameter monitoring: The immersion solution was replaced every month, and the test pieces were taken out every three months to measure the mass, chloride ion concentration, and penetration depth.

[0017] Furthermore, the wind tunnel experiment comprises the following steps:

[0018] (1) Experimental device: a wind tunnel test chamber with temperature, salt, humidity and dynamic multi-field coupling is used;

[0019] (2) Temperature gradient: Three groups of temperature conditions were set: normal temperature reference group (25°C), high temperature group (40°C) and extremely high temperature group (55°C);

[0020] (3) Salinity gradient: the chloride concentration of the solution was set at 20‰, 25‰ and 30‰;

[0021] (4) Humidity and zone division: The water level is set at 40 cm, and the height is divided into the following three sections: below 20 cm is the full immersion zone;

[0022] 20-40cm is the dry-wet cycle zone;

[0023] 40-60cm is the atmospheric zone;

[0024] (5) Dry-wet cycle: water is automatically pumped out every 12 hours, with a pumping height of 20 cm;

[0025] (6) Wave-making flushing: Set up a wave-making pump for 24-hour flushing;

[0026] (7) Sample specifications: The specimen size is 20*20*60cm (length, width and height);

[0027] (8) Sampling frequency: 4 times a year, 3 samples each time, a total of 12 samples, 3 samples are kept as backup;

[0028] (9) Monitoring indicators: The immersion solution was replaced monthly, and the mass, chloride ion concentration, and penetration depth of the samples were measured every 3 months.

[0029] Furthermore, the detection method is a fusion of Attention U-Net and CNN.

[0030] Furthermore, the fusion strategy is any one of multi-level feature extraction, attention mechanism in skip connection and joint loss function.

[0031] Furthermore, the basic framework of the Attention U-Net is: encoder-decoder architecture and attention module.

[0032] Furthermore, the core components of the CNN include convolutional layers, pooling layers and activation functions.

[0033] Beneficial effects of the present invention:

[0034] 1) The influence of the marine environment on all factors of marine concrete is fully considered, and the corrosion degradation similarity theory between model experiments and prototypes is vigorously developed, providing support for the laboratory accelerated test method of marine concrete.

[0035] 2) It solves the problem of scarce samples due to long-term concrete deterioration and low accuracy of machine vision methods due to insufficient learning samples.

[0036] 3) Study the influence of marine environmental corrosion and dynamic effects on the durability of concrete structures

[0037] 4) Establish a concrete life prediction model that takes into account high-performance new materials and marine environmental characteristic parameters BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a diagram showing the mechanism of reduced durability of concrete structures due to chemical corrosion and physical deterioration;

[0039] Figure 2 This is the framework diagram of CNN and Attention U-Net network;

[0040] Figure 3 This is the structure diagram of the attention gate. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below in conjunction with specific embodiments so that those skilled in the art can understand.

[0042] Example 1 Coupling experiment

[0043] 1.1 Ocean immersion experiment

[0044] 1) Sample placement: The specimens are placed in a specific area on the seashore, which is divided into full immersion area, splash zone and atmospheric zone.

[0045] 2) Quantity configuration: 5 samples are placed in each area, for a total of 15.

[0046] 3) Regular sampling: Sampling is conducted 4 times a year, 3 samples are taken each time, for a total of 12 samples, and 3 samples are reserved as spare samples.

[0047] 4) Parameter monitoring: Replace the immersion solution every month, take out the test pieces every 3 months, and measure the mass, chloride ion concentration and penetration depth.

[0048] 1.2 Wind Tunnel Test Box

[0049] 1) Experimental device: A wind tunnel test chamber with temperature, salt, humidity and dynamic multi-field coupling is used.

[0050] 2) Temperature gradient: Set three groups of temperature conditions - normal temperature reference group (25℃), high temperature group (40℃) and extremely high temperature group (55℃).

[0051] 3) Salinity gradient: The chloride concentration of the solution was set at 20‰, 25‰ and 30‰.

[0052] 4) Humidity and area division: The water level is set at 40cm, and the height is divided into the following three sections:

[0053] The area below 20cm is the full immersion area

[0054] 20-40cm is the dry and wet cycle area

[0055] 40-60cm is the atmospheric zone

[0056] 5) Wet-dry cycle: water is automatically pumped out every 12 hours, with a pumping height of 20cm.

[0057] 6) Wave-making flushing: Set up a wave-making pump for 24-hour flushing.

[0058] 7) Sample specifications: The specimen size is 20*20*60cm (length, width and height).

[0059] 8) Sampling frequency: Sampling is conducted 4 times a year, 3 samples are taken each time, for a total of 12 samples, and 3 samples are retained as backup.

[0060] 9) Monitoring indicators: Replace the soaking solution every month and measure the sample quality, chloride ion concentration and penetration depth every 3 months.

[0061] Example 2 Improved Attention U-Net Deep Fully Convolutional Network Machine Vision Detection Method

[0062] 2.1 Attention U-Net network

[0063] Attention U-Net is an extension of the U-Net network. It introduces the attention mechanism to enhance the ability to capture the target area. The basic structure of the Attention U-Net network is as follows:

[0064] 1) Encoder-Decoder Architecture: Similar to U-Net, Attention U-Net uses an encoder-decoder structure to extract low-level features of the image through convolution and downsampling, and then reconstruct the features into outputs with the same resolution as the input through the decoder.

[0065] 2) Attention module: In the skip connections of U-Net, Attention U-Net introduces an attention module to filter out irrelevant background information and retain only features related to the target, thereby improving the segmentation accuracy of the model.

[0066] The output calculation formula of the attention module is:

[0067] α=σ(W f ·f+W g ·g+b)

[0068] in:

[0069] f and g are feature maps,

[0070] W f and W g is the learned weight matrix,

[0071] b is the bias,

[0072] σ represents the Sigmoid activation function, which controls the attention weight.

[0073] After the Attention mechanism calculates the attention weight α, it applies it to the feature map to strengthen the target area:

[0074] f′=α·f

[0075] Among them, f′ is the enhanced feature map

[0076] 2.2 Convolutional Neural Network (CNN)

[0077] Convolutional neural networks extract local features through hierarchical convolution and pooling operations and are widely used in image classification and detection tasks. The core components of CNN include:

[0078] 1) Convolution layer: The input image or feature map is scanned through the convolution kernel to extract local features in space. The formula is:

[0079]

[0080] Among them, x is the input image, w is the convolution kernel weight, and b is the bias

[0081] 2) Pooling layer: used for downsampling, reducing the size of feature maps while retaining key information. Commonly used pooling methods include maximum pooling and average pooling.

[0082] 3) Activation function: The ReLU function, i.e. f(x) = max(0, x), is usually used to perform a nonlinear transformation on the convolution result.

[0083] 2.3 Method of integrating Attention U-Net and CNN

[0084] In order to fuse CNN and Attention U-Net, CNN can be introduced in the encoding stage to extract more local features, and the Attention module can be added in the decoding stage to enhance the feature attention ability. The following are some commonly used fusion strategies:

[0085] 1) Multi-level feature extraction: CNN is used on each layer of the encoder to extract multi-scale features and input them into the attention module to better capture the details of the image.

[0086] 2) Attention mechanism in skip connection: An attention module is added to the skip connection to selectively transfer features by adjusting the weights, thereby improving the segmentation effect.

[0087] 3) Joint loss function: A joint loss function is used to optimize the segmentation and feature extraction performance of the network.

[0088] By streamlining the attention gate module, a full convolutional network is constructed on the Attention U-Net network. The overall framework of Attention U-Net is as follows: Figure 2 shown.

[0089] The deconvolution layer upsamples the low-resolution feature map through transposed convolution. The kernel size of the transposed convolution is 2×2, and the stride is 2. In each transposed convolution operation, each value of the low-resolution input feature is multiplied by all the weights in the convolution kernel, and then these results are mapped to the corresponding position of the high-resolution output. Next, through the concatenation operation, the feature map extracted from the jump connection is merged with the upsampled feature map to make full use of the low-level and high-level features. The input size of the concatenation operation in this article is the same (H×W×C), and the number of output channels is doubled after concatenation to (H×W×2C). The attention gate adopts a self-attention mechanism to calculate the attention coefficient through global features, thereby filtering out irrelevant parts of the local features. To simplify the calculation, the gating vector of each attention gate is directly selected from the features output by the deconvolution layer to keep the same size as the input feature map, thereby eliminating the upsampling operation of the attention gate. The structure of the attention gate is as follows Figure 3 shown.

[0090] Although the above embodiments have been described in detail, they are only a part of the embodiments of the present invention, not all of them. People can also obtain other embodiments based on this embodiment without creativity, and these embodiments all belong to the protection scope of the present invention.

Claims

1. A method for detecting the durability of marine concrete under multi-field coupling conditions using machine vision, characterized in that: It includes ocean immersion experiments and wind tunnel experiments, which respectively simulate the corrosion degradation process in the marine environment and verify each other.

2. The detection method according to claim 1, characterized in that: The ocean immersion experiment comprises the following steps: (1) Sample placement: The specimens are placed in a specific area on the seashore, which is divided into a full immersion area, a splash zone, and an atmospheric zone; (2) Quantity configuration: 5 samples are placed in each area, for a total of 15; (3) Regular sampling: 4 times a year, 3 samples each time, a total of 12 samples, 3 samples reserved as spare samples; (4) Parameter monitoring: The immersion solution was replaced every month, and the test pieces were taken out every three months to measure the mass, chloride ion concentration, and penetration depth.

3. The detection method according to claim 1, characterized in that: The wind tunnel experiment comprises the following steps: (1) Experimental device: a wind tunnel test chamber with temperature, salt, humidity and dynamic multi-field coupling is used; (2) Temperature gradient: Three groups of temperature conditions were set: normal temperature reference group 25°C, high temperature group 40°C, and extremely high temperature group 55°C; (3) Salinity gradient: the chloride concentration of the solution was set at 20‰, 25‰ and 30‰; (4) Humidity and zone division: The water level is set at 40 cm, and the height is divided into the following three sections: below 20 cm is the full immersion zone; 20-40cm is the dry-wet cycle zone; 40-60cm is the atmospheric zone; (5) Dry-wet cycle: water is automatically pumped out every 12 hours, with a pumping height of 20 cm; (6) Wave-making flushing: Set up a wave-making pump for 24-hour flushing; (7) Sample specifications: The specimen size is 20*20*60cm in length, width and height; (8) Sampling frequency: 4 times a year, 3 samples each time, a total of 12 samples, 3 samples are kept as backup; (9) Monitoring indicators: The immersion solution was replaced monthly, and the mass, chloride ion concentration, and penetration depth of the samples were measured every 3 months.

4. The detection method according to claim 1, characterized in that: The detection method is a fusion of Attention U-Net and CNN.

5. The detection method according to claim 4, characterized in that: The fusion strategy is any one of multi-level feature extraction, attention mechanism in skip connection and joint loss function.

6. The detection method according to claim 4, characterized in that: The basic framework of the Attention U-Net is: encoder-decoder architecture and attention module.

7. The detection method according to claim 4, characterized in that: The core components of the CNN include convolutional layers, pooling layers and activation functions.

Images