Distributed Internet of Things system for mass concrete construction quality monitoring

Abstract

The invention discloses a distributed Internet of Things system for mass concrete construction quality monitoring, which can realize high-precision data acquisition of mass concrete temperature strain and perform effective crack monitoring in the later period of structure maintenance, and at the same time, an intelligent algorithm is embedded in edge intelligent gateway equipment, so that the construction quality of the mass concrete is monitored. And the construction quality of mass concrete is guaranteed. The system comprises an edge end and a cloud end, the edge end comprises an online crack monitoring camera, an edge intelligent gateway, a temperature sensor and a strain sensor, and the cloud end is equipped with a computing server. According to the invention, an edge intelligent gateway device, an online crack monitoring camera and a cloud edge cooperation mechanism based on multi-level edge calculation are established, a novel attention network structure with multi-level and multi-scale feature fusion is provided, and more accurate crack segmentation is realized. And a set of simple and efficient crack quantitative characterization algorithm is provided on the basis of the fine crack segmentation image, the actual size information of the crack is accurately quantified, and the real-time condition of the crack is mastered.

Description

Technical Field

[0001] The present invention belongs to the field of concrete monitoring technology, and specifically relates to a distributed Internet of Things system for monitoring the quality of large-volume concrete construction. Background Art

[0002] When certain dimensions of a concrete structure reach certain standards and may cause significant temperature changes and shrinkage due to hydration heat, the concrete structure is considered to be mass concrete. Due to its large volume and mass, the hydration heat will produce local temperature rise inside. If the temperature rises too quickly, it will lead to significant temperature gradients and differences, increase temperature stress, and then produce temperature difference cracks, reduce structural safety and reliability, and cause significant losses. Therefore, during the construction of mass concrete, it is necessary to monitor the temperature and strain of the entire process, and combine real-time monitoring data to prompt the construction site to take corresponding maintenance measures to prevent the occurrence of harmful cracks. However, traditional manual detection methods are difficult to achieve the management requirements of large-scale pouring. The monitoring equipment has high cost and energy consumption, complex operation, and large size.

[0003] Crack detection of large-volume concrete in the later stage of maintenance is a crucial link, which is directly related to the safety and durability of the structure. However, the currently commonly used manual visual inspection method has many limitations, including insufficient recognition of small cracks, low detection efficiency, high cost, and susceptibility to environmental factors. Concrete crack detection technology based on digital image processing has become a research hotspot because it can overcome the problems of time-consuming and labor-intensive traditional manual detection and insufficient detection accuracy. However, the background and characteristics of the cracks are complex, and the existing image algorithms often require manual intervention, with a low degree of automation, making it difficult to conduct continuous real-time monitoring.

[0004] In general, it is important to effectively monitor the development of temperature and strain in real time and to identify the occurrence of cracks in a timely and accurate manner during the construction and maintenance of large-volume concrete. However, various current monitoring methods still have defects, specifically: (1) The current large-volume concrete temperature and strain monitoring methods are costly and energy-intensive, inefficient, and difficult to meet the real-time management needs of large-area concrete pouring; (2) Post-molding concrete crack detection is mainly manual, time-consuming and labor-intensive, and has insufficient detection accuracy. Existing image algorithms require manual intervention, are inefficient, and are difficult to perform continuous real-time monitoring, lacking automated methods. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a distributed Internet of Things system for monitoring the construction quality of large-volume concrete. The system aims to realize high-precision data collection of temperature strain during the construction of large-volume concrete and perform effective crack detection in the later stage of structure maintenance. The system realizes real-time analysis of various data and unified scheduling management of different equipment through multi-level edge computing and cloud-edge collaboration, thereby ensuring the construction quality of large-volume concrete.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A distributed Internet of Things system for mass concrete construction quality monitoring, comprising: an edge end and a cloud end, wherein the edge end comprises an online crack monitoring camera, an edge intelligent gateway, a temperature sensor and a strain sensor, and the cloud end is equipped with a computing server;

[0008] The online crack monitoring camera includes an image acquisition module, an image processing module and a data transmission module; the image acquisition module is used to capture images of the concrete structure, the image processing module has an embedded crack recognition network and a crack quantitative characterization unit, which are used to identify cracks in the image and calculate crack information, respectively, and the data transmission module outputs the crack information to a computing server in the cloud;

[0009] The temperature sensor and the strain sensor are arranged on the concrete structure, and are used to detect the temperature data and the strain data of the concrete structure respectively; the edge intelligent gateway transmits the temperature data and the strain data to the computing server in the cloud;

[0010] The computing server performs computational analysis on the crack information, temperature data, and strain data, and sends instructions to the online crack monitoring camera and the edge intelligent gateway.

[0011] Optionally, the embedded crack identification network is based on U-Net, and a normalization layer and an activation function are not simultaneously attached after a convolution layer, and only one of the normalization layer and the activation function is attached after the same convolution layer.

[0012] Optionally, the standardization layer adopts the following layer standardization method:

[0013]

[0014] In the formula, The calculation result of the representation layer standardization, x i represents the input feature map, i represents the index of the input feature map, and Respectively represent the mean and standard deviation of the data set currently fed into the normalization layer;

[0015] The activation function adopts the following H-Swish function:

[0016]

[0017] In the formula, a is the input of the activation function, ReLU6 is the ReLU function after maximum suppression,

[0018] Optionally, the specific process of the crack quantitative characterization unit calculating the crack information is as follows:

[0019] Perform threshold segmentation on the crack recognition results of the image to determine whether each pixel in the image is a crack;

[0020] Cracks are divided according to the connectivity between pixels, and the total number of cracks is counted;

[0021] Calculate the total area of ​​the cracks;

[0022] Extract the crack skeleton and calculate the total length of the crack;

[0023] Calculate the average width of the cracks;

[0024] The total area, total length and average width of the cracks are mapped from the pixel space to the physical space to obtain the actual total area, total length and average width of the cracks.

[0025] Optionally, the crack recognition result of the image is subjected to threshold segmentation to determine whether each pixel in the image is a crack, specifically:

[0026] The binary classification threshold is selected as 0.5, and the crack pixel-level recognition results are binarized as follows:

[0027]

[0028] In the formula, Pr t represents the prediction result of the model for the tth image, T represents the total number of images in the predicted image time series set, and Pr t (i, j) = 1 means that the pixel at position (i, j) in the image is predicted to be a crack, Pr t (i, j)=0 indicates that the position (i, j) is predicted to be a non-crack, and H and W represent the spatial height dimension and the spatial width dimension, respectively.

[0029] Optionally, the division of cracks according to connectivity between pixels is specifically as follows:

[0030] Starting from any crack pixel, all crack pixels connected to it are found according to the crack connectivity criterion, and the total connected domain obtained represents a crack; after visiting all crack pixels, all cracks in the image are divided;

[0031] The connectivity criterion of the crack is: if one pixel is a crack pixel and there are crack pixels in the surrounding eight pixels, then these crack pixels are interconnected; in one crack, all pixels are interconnected.

[0032] Optionally, the total area of ​​all cracks is calculated using the following formula:

[0033]

[0034] In the formula, Represents the total crack area of ​​the tth image Represents the result of binarization, and s represents the area of ​​a pixel in the image.

[0035] Optionally, the extracting of each crack skeleton and calculating the total length of the cracks is specifically as follows:

[0036] The single-pixel-width skeleton of the crack is extracted, and the total length of the crack is calculated by the following formula:

[0037]

[0038] In the formula, represents the total length of cracks in the tth image, Sk t Represents the skeleton graph of the tth image, and l represents the side length of a pixel in the image.

[0039] Optionally, the average width of the crack is calculated using the following formula:

[0040]

[0041] In the formula, represents the average crack width of the tth image.

[0042] Optionally, the total area, total length and average width of the cracks are mapped from the pixel space to the physical space to obtain the actual total area, total length and average width of the cracks, specifically:

[0043] Calibrate the online crack monitoring camera and calculate the proportionality coefficient ψ between the actual physical size and the image pixel;

[0044] The actual total area, total length and average width of the cracks are calculated by the following formula:

[0045]

[0046] In the formula, and They represent the actual total area, total length and average width of the cracks respectively.

[0047] The beneficial effects of the present invention are:

[0048] (1) In order to solve the problems of traditional large-volume concrete temperature strain monitoring methods being complex, inefficient, high hardware system cost and energy consumption, and unable to conduct real-time monitoring and analysis, the present invention has developed an edge intelligent gateway device based on embedded technology. The device can collect and wirelessly transmit multi-type and multi-interface sensor signals according to monitoring requirements, and can pre-process abnormal data with edge computing capabilities.

[0049] (2) In view of the problems of low detection efficiency, high cost, susceptibility to environmental factors, and low degree of automation in existing concrete crack monitoring methods, the present invention has developed an online crack monitoring camera with an embedded optimized crack precision identification and quantification algorithm. It can directly output information such as crack length and area through edge computing capabilities, and transmit data through wireless communication capabilities.

[0050] (3) In order to reduce the transmission pressure at the edge and ensure the analysis rate, the present invention proposes a lightweight segmentation attention network model CrackU-Net. This network structure breaks through the two-level fusion design paradigm and achieves high-precision recognition of cracks and accurate capture of subtle changes by integrating multi-level features and multi-scale features. At the same time, the deep learning model has low computational complexity and few parameters, and can still maintain its original high performance on edge devices with limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is an architectural diagram of a distributed Internet of Things system for mass concrete construction quality monitoring.

[0052] Figure 2 It is a crack detection flow chart.

[0053] Figure 3 It is a monitoring area map.

[0054] Figure 4 This is the temperature measurement point arrangement diagram.

[0055] Figure 5 It is the arrangement diagram of strain measuring points.

[0056] Figure 6 It is the overall monitoring layout diagram.

[0057] Figures 7a to 7d It is the temperature monitoring result diagram, which are the temperature curves of point 1, point 2, point 3 and point 4 respectively.

[0058] Figures 8a to 8d It is the strain monitoring result diagram, which are the strain curves of point 1, point 2, point 3 and point 4 respectively.

[0059] Fig. 9This is the effect diagram of crack pixel-level division based on connectivity. DETAILED DESCRIPTION

[0060] The present invention will now be described in further detail with reference to the accompanying drawings.

[0061] The present invention proposes a distributed Internet of Things system for mass concrete construction quality monitoring. The system adopts the architecture of "distributed collection at the edge + centralized recycling and processing at the cloud end". Figure 1 As shown in the figure, the system construction mainly includes the following two aspects:

[0062] First, establish an IoT sensing and collection integrated system for large-volume concrete construction quality monitoring, including edge intelligent gateway equipment, online crack monitoring cameras, temperature sensors, strain sensors, and a cloud-edge collaborative mechanism based on multi-level edge computing;

[0063] Second, a new attention network structure CrackU-Net with multi-level and multi-scale feature fusion is proposed, which promotes the thorough flow and fusion of global and local information, strengthens the learning task of small-proportion crack targets, and achieves more accurate crack segmentation; based on the fine crack segmentation image, a simple and efficient crack quantitative characterization algorithm is proposed to accurately quantify the actual size information of the crack and grasp the real-time status of the crack.

[0064] For the first aspect, the distributed IoT system for mass concrete construction quality monitoring includes the edge and cloud.

[0065] The edge includes online crack monitoring cameras with edge computing capabilities, edge intelligent gateways, temperature sensors and strain sensors. The edge intelligent gateways and online crack monitoring cameras can be deployed in a distributed manner thanks to their wireless communication and edge computing capabilities.

[0066] The edge intelligent gateway includes an acquisition module, a communication module and a computing module, which can collect temperature, strain, and multi-type, multi-interface sensor signals, pre-process abnormal data, and wirelessly transmit them.

[0067] The online crack monitoring camera includes an image acquisition module, an image processing module and a data transmission module. The camera has an embedded optimized high-precision crack recognition and quantification algorithm, which can process images in real time through edge computing and directly output crack information.

[0068] The cloud is equipped with high-performance computing servers. The edge can process most of the data and execute the decisions and control operations from managers locally, thereby reducing the delay caused by data transmission. The cloud can provide support for overall large-scale data analysis and maintenance decisions.

[0069] For the second aspect, the present invention proposes a new attention network structure CrackU-Net with multi-level and multi-scale feature fusion, which promotes the thorough flow and fusion of global and local information, while strengthening the learning task of small-proportion crack targets to achieve more accurate crack segmentation; based on the fine crack segmentation image, a set of simple and efficient crack quantitative characterization algorithms is proposed to accurately quantify the actual size information of the cracks and grasp the real-time status of the cracks.

[0070] U-shaped structure networks represented by U-Net have made significant progress in the field of biomedical image segmentation. Its unique U-shaped structure design is an effective feature fusion mechanism that can achieve significant segmentation performance. Inspired by the U-shaped structure, the present invention proposes CrackU-Net. It innovates the two-level fusion design paradigm of the traditional U-shaped structure network and proposes the idea of ​​multi-level fusion, which can further enhance the information flow within the network and achieve more effective and thorough multi-level and multi-scale feature fusion. The network is mainly composed of 5 encoder modules, 4 upsampling blocks, 3 multi-scale feature aggregation modules, 1 side output block and 1 final output block, such as Figure 2 As shown in the figure. Thanks to this multi-level feature fusion design, CrackU-Net can integrate low-level detail information into high-level features with richer semantic information but less details, reduce the information loss in the feature transmission process, and enable the network to capture high-level semantic features and low-level fine-grained information at the same time, effectively improving the accuracy and precision of image segmentation. The latest research on convolutional neural networks has empirically proved that using fewer activation functions and normalization layers and more reasonable arrangements of their embedding positions and numbers will significantly improve the expressiveness of the model. Therefore, the normalization layer and activation function of the model are rearranged. For most modern convolutional neural network structures, the usual practice is to attach the activation function to the normalization layer, and then stack them uniformly behind a convolution layer, that is, to form a triple structure of "convolution → normalization layer → activation function". However, in CrackU-Net, a different design is adopted. For all five stages, the normalization layer and activation function are not attached to the convolution layer at the same time. Only one of them will be attached to the same convolution layer to reduce the use of both. Experimental results show that this is a better design than the commonly used triple structure and can improve model performance. The design and modification of the normalization layer and activation function are as follows:

[0071] The present invention adopts the layer normalization method to replace the conventional batch normalization method. Assume that the input feature map to be calculated by layer normalization is Then the calculation process of layer normalization can be described as

[0072]

[0073] In the above formula, i represents the index of the input feature map. Since the research of the present invention is centered around two-dimensional images, i is a four-dimensional vector: i = (i N ,i C ,i H ,i W ), where N represents the Batch axis, C represents the channel axis, H refers to the spatial height dimension, and W refers to the spatial width dimension. Representation layer Normalization layer, Represents the average value of the data set of the current feeding layer standardization layer, which can be calculated by the following formula:

[0074]

[0075] in, The set of indices that represent the data to be standardized, and k represents Any element in , obviously k is also a four-dimensional vector. Therefore, Corresponding to the data set, it is a subset of the input feature map pixels, with x k This corresponds to the pixel with subscript index k in the current number set. n represents the set The total number of elements in . Next, calculate the standard deviation of the current set of numbers by the following method

[0076]

[0077] Here, ∈ is a small constant to avoid the denominator being equal to zero. Different normalization methods have different Division criteria, in the layer standardization method, each sample is divided into a The mean and standard deviation (μ, σ) are calculated within the sample, that is, in the input feature map x, the pixels sharing the same N index are normalized together. In other words, (μ, σ) in the layer normalization method is calculated along the direction of (C, H, W). Therefore, the layer normalization method It can be defined by the following formula:

[0078]

[0079] In the above formula, k N is the sub-index of the four-dimensional vector k along the N-axis, i N is the sub-index of the four-dimensional vector i along the N-axis. The layer normalization method in these algorithms uses another set of methods different from LN_He (a layer normalization method). Partitioning rules, specifically, at this time (μ, σ) is only calculated along the (c) direction, that is, the set The data in all share the same N, H, and W indexes:

[0080]

[0081] In the formula, k H (and i H ) and k W (and i W ) represent the sub-index of k (and i) along the H-axis and W-axis respectively. In order to compensate for the loss of representation ability that may be caused by the normalization of values, LN_He and LN_VT (another layer normalization method) add learnable affine transformation parameters γ and β to each channel of the feature map:

[0082]

[0083] Where yi represents the final output of the LN_He or LN_VT method, and Represented as each channel i c The added scaling and translation parameters (also called weights and biases) are equivalent to performing a linear transformation.

[0084] For the activation function, the H-Swish function is introduced into CrackU-Net, which can be defined as follows:

[0085]

[0086] Among them, a is the input of the activation function, and ReLU6 is a maximum-suppressed version of the ordinary ReLU function, which limits the output after activation to less than or equal to 6. This is very effective for low-precision mobile devices and avoids the precision loss that may be caused by ReLU. The specific calculation formula of ReLU6 is as follows:

[0087]

[0088] In addition, GELU, a smoother variant of ReLU, has been used in the most advanced Transformer methods and has brought different degrees of performance improvements to these algorithms. The calculation formula of GELU is

[0089]

[0090] In the formula, b is the input of the GELU function, Φ(x) is the cumulative distribution function of the standard normal distribution. Since it is an S-type function, it is usually approximated by the Tanh function. P(X≤b) represents the probability that the random variable X is less than b, and erf is the error function. Because GELU is a better alternative function than ReLU, the ReLU activation function is replaced by GELU in the shallow network layer of CrackU-Net.

[0091] After being processed by the above algorithm, the steps for calculating the quantitative information related to cracks are as follows:

[0092] (1) Threshold segmentation of crack pixel-level identification results

[0093] The semantic segmentation network described above maps the input disease images to the range of (0, 1), and the value reflects the possibility that the pixel is a crack. In order to facilitate subsequent processing, the predicted image is binarized according to a certain threshold. The selected binary classification threshold is 0.5.

[0094]

[0095] In the formula, Pr t Pr represents the prediction result of the model for the tth image, and T represents the total number of images in the predicted image time series set. t (i, j) = 1 means that the pixel at position (i, j) in the image is predicted to be a crack, Pr t (i, j) = 0 indicates that the position (i, j) is predicted to be a non-crack (ie, a background pixel). H and W represent the spatial height dimension and the spatial width dimension, respectively.

[0096] (2) Count the total number of cracks

[0097] In most cases, there is not just one crack in the image, but multiple cracks. Obviously, the object detection method at the bounding box level cannot accurately divide the cracks, and there will be many situations such as overlapping and misjudgment of the bounding boxes. The segmentation method based on the present invention can obtain the crack pixel level representation Pr t (i, j), and then the cracks can be accurately divided based on the connectivity between pixels. The comparison between crack division based on connectivity and target detection is shown in Figure 8.

[0098] There are two criteria for the connectivity of cracks: (a) If a pixel is a crack pixel and there are crack pixels in the surrounding 8 pixels, these crack pixels can be considered to be interconnected; (b) In a crack, all pixels are interconnected. Based on these two criteria, we can start from any crack pixel and find all the crack pixels that are interconnected with it. This total connected domain represents a crack. After visiting all the crack pixels, all the cracks in the image are divided.

[0099] (3) Calculation of total crack area

[0100] The total area of ​​the crack It can be calculated by the following formula:

[0101]

[0102] Where G(x, y) is the geometric correction coefficient, which is used to calibrate the pixel displacement in the recognition result. ds is the area of ​​the crack, which can be roughly considered as the area of ​​a pixel. The present invention assumes that the captured image has no geometric deformation such as distortion, so the calibration coefficient is simplified to G(x, y)≡1, so the total area of ​​the crack is approximately equal to the area of ​​all crack pixels. After binarizing the prediction results, since the value of non-crack pixels is 0, the total area of ​​cracks in the tth image is It can be calculated by the following formula:

[0103]

[0104] Where s represents the area of ​​a pixel in the image. Represents the result of binarization.

[0105] (4) Calculation of crack skeleton extraction and total length

[0106] After using the skeletonization method to extract the crack skeleton with a single pixel width, the length of the crack can be calculated as

[0107]

[0108] In the formula, dl represents the length of the crack skeleton line. It can be roughly assumed that the length of the skeleton line is the length of the skeleton pixel, so the total length of the crack is approximately equivalent to the total length of the crack skeleton pixels. Let the skeleton image of the tth image be Sk t , since the skeleton image is also a {0, 1} binary image, the crack length of the tth image is It can be calculated by the following formula:

[0109]

[0110] Wherein, l represents the length or width of a pixel in the image (the pixels of the image in the present invention are considered to be square, with an aspect ratio of 1:1, and the length and width are equal).

[0111] (5) Calculation of average crack width

[0112] According to the total crack area and total crack length calculated by the above method, the average width of the crack can be obtained:

[0113]

[0114] In the above formula, represents the average crack width of the tth image.

[0115] (6) Mapping from pixel space to physical space

[0116] The above steps (3) to (5) are all about solving the crack parameters at the pixel level, all in pixels. However, in actual engineering, we are more concerned about the actual physical size of the crack in the real world to evaluate the working condition of the structure. By calibrating the camera, we can calculate the proportional coefficient ψ between the actual physical size and the image pixel, and then based on ψ, we can get the number of defective pixels in the image. The actual physical size of the disease The mapping relationship between them is:

[0117]

[0118] Therefore, the total area of ​​the crack in the real world can be obtained Overall length and the average width

[0119]

[0120]

[0121] Next, the specific implementation steps of the present invention are described with reference to examples.

[0122] (1) System deployment plan

[0123] The main body of the raft foundation of a nuclear power plant is a cylindrical structure with a radius of 27m and a thickness of 3.70m. There is a concrete prestressed tensioning corridor at the bottom of the raft foundation. The corridor is a circular ring structure, concentric with the raft foundation, with a radius of R=20.68m. The wall thickness on both sides of the corridor is 0.8m, the corridor width is 2.5m, and the top is a concrete cover plate. Crack control is an important quality measure for large-volume concrete construction in nuclear power plants. During the pouring and curing process, the development and heat dissipation process of hydration heat of raft foundation concrete produces complex temperature and stress fields in the raft foundation. The temperature and temperature stress are difficult to control, and the potential risk of cracking is high. For this reason, it is necessary to carry out temperature and strain monitoring during the pouring and curing of concrete, and to achieve overall control of concrete temperature and temperature stress through dynamic curing measures, so as to ensure the molding quality of concrete. The large-volume concrete monitoring this time mainly includes three aspects of temperature, strain and crack monitoring. The monitoring area is located in 1 / 4 of the raft foundation. The monitoring area is as follows: Figure 3 shown.

[0124] (2) Data Collection

[0125] Arrangement of temperature measuring points: According to the specifications and finite element calculation results, the temperature sensors are mainly arranged in two radial directions of 0° (BB), 45° (DD) and 90° (AA). Three points are selected in the 0° direction, namely 6.5m from the center (1#), 13.5m (2#), and 22m (3#); three points are selected along the 90° radius direction, namely 6.5m from the center (4#), 22m (5#) and radius point (6#); two points are selected along the 45° radius direction, namely 13.5m from the center (7#) and radius point (8#); select the center point to arrange the sensor (9#); 1 atmospheric temperature measuring point; 2 insulation layer temperature measuring points; 2 insulation shed temperature measuring points; the overall design of the raft foundation temperature measuring point arrangement is as follows Figure 4 shown.

[0126] Strain measurement point arrangement: According to the pouring direction of raft foundation concrete, strain sensors are mainly arranged in two radial directions of 0° (BB) and 90° (AA). Three points are selected in the 0° direction, namely 6.5m from the center (1#), 13.5m (2#), and 22m (3#); three points are selected along the 90° radial direction, namely 6.5m from the center (4#), 22m (5#), and radius point (6#); two points are selected along the 45° radial direction, namely 13.5m from the center (7#) and radius point (8#); the overall design of raft foundation strain measurement point arrangement is as follows Figure 5 shown.

[0127] Crack measurement point arrangement: combined with finite element analysis results and previous monitoring data, 1) high stress area in finite element calculation; 2) abnormal cooling rate area in temperature monitoring; 3) strain peak area in strain monitoring; 4) characteristic points and edge areas.

[0128] Overall layout of the inspection: Use intelligent gateway equipment that supports 8-channel temperature and 8-channel strain synchronous acquisition and online crack monitoring camera equipment, and determine the location of the equipment according to the temperature measurement points and strain measurement points. Set up a monitoring room 800m outside the edge of the raft foundation in the 0° direction (the location of the monitoring room is adjusted according to the actual situation on site and is set in the area outside the edge of the raft foundation). Each wireless acquisition module is connected to the sensor at the corresponding position through a data cable, and the data is aggregated to the host in the monitoring room through wireless transmission. The layout diagram is shown in the figure below. Figure 6 shown.

[0129] (3) Test results analysis

[0130] Temperature monitoring results: 1) Maximum temperature: The maximum temperature of the raft foundation concrete reached 73.96°C, which was observed 95.5 hours after pouring; 2) Temperature distribution characteristics: The peak temperatures of the three layers of concrete at point 1# were 53.91°C, 72.38°C, and 69.38°C respectively. During the heating stage, the temperature of the middle layer was generally higher, followed by the upper layer, and the temperature of the lower layer was relatively low; 3) Characteristics of the heating stage: The heating stage of the bottom and side walls was slightly shorter, and the heating stage of the center was longer. During the cooling stage, the average cooling rates of the upper and middle layers are similar, while the cooling rate of the lower layer is smaller; 4) Bedrock insulation effect: As time goes by, the heat of the raft foundation concrete gradually dissipates, the temperature difference of the contact surface with the bedrock gradually decreases, and the temperature reduction rate of the lower layer gradually decreases, presenting a smooth cooling curve, indicating that the bedrock has a good insulation effect; 5) Temperature field distribution law: In the vertical comparison, the middle temperature is higher and the upper and lower layers are lower; in the radial comparison, the symmetrical position of the temperature in the same layer is basically the same, and gradually decreases along the radial direction; in the circumferential comparison, the temperature in the same layer and the same radius area is basically the same. 6) Cooling rate: The cooling rate of all measuring points every 24 hours is maintained below 2.0℃ / d; 7) Inside and outside temperature difference: The measured values ​​of the inside and outside temperature difference are all controlled within 25℃, meeting the requirements of the specification. Some results of temperature monitoring are as follows Figures 7a to 7d shown.

[0131] Strain monitoring results: When concrete shrinkage is not considered, the circumferential and radial strains of the upper, middle and bottom layers in the center of the raft foundation are all compressive strains. After considering shrinkage, the circumferential and radial strains of the bottom and middle layers are still compressive strains, and the top layer is also basically compressive strain. During the curing period, the radial and circumferential strains of the upper layer are compressive stresses, and some of them are tensile stresses initially, which turn into compressive stresses with the hydration heat and cooling. The radial and circumferential strains of the middle layer are compressive strains, and the strains decrease as the solidification period is extended. The radial and circumferential strains of the lower layer are tensile stresses initially, which turn into compressive stresses with the hydration heat and cooling. The strain monitoring results are as follows Figures 8a to 8d shown.

[0132] Analysis of crack results: Comparing the performance of the proposed algorithm with different segmentation algorithms, as shown in Table 1, CrackU-B has achieved a more excellent disease recognition effect on the basis of far fewer parameters and computational complexity than other comparison models (it should be stated in advance that since the model proposed by Ni et al. is based on the TensorFlow framework, the ptflops tool cannot be used, so in order to reflect the fairness of the comparison, MACs are not reported for these two networks). Taking the U-Net model as an example, the number of parameters of CrackU-B is 41.06% of that of U-Net, and the computational complexity is only 20.38%, but the evaluation indicators mi IoU and mi Dice are 6.39% and 3.98% higher than U-Net, respectively. In addition, the number of parameters and computational complexity of CrackU-M and CrackU-L are slightly increased compared with CrackU-B, but these two models can obtain more accurate recognition effects. In order to verify the effectiveness of the proposed nonlinear activation function combination, different activation function combinations are considered and embedded in the model, and the corresponding running results are shown in Table 2. Obviously, compared with other design schemes, the proposed GELU+H-Swish combination can obtain the highest mi IoU and mi Dice. The impact of batch normalization method and different layer normalization methods on model prediction performance are considered, and the corresponding model recognition results are shown in Table 3. The proposed LN_VT layer normalization strategy can obtain the highest mi IoU and mi Dice.

[0133] Table 1 Performance comparison of different segmentation algorithms

[0134] Model Name mi IoU(%) mi Dice(%) #Param.(M) MACs(G) U-Net 90.54 94.38 7.77 26.00 U-Net(1arge) 96.02 97.69 31.04 103.52 U-Net4+ 86.24 91.56 36.63 261.22 Attention U-Net 95.94 97.63 34.88 125.98 4Dense(6666) 90.95 94.62 4.16 / 4Dense(7777) 93.41 96.23 5.02 / CrackU-B 96.93 98.36 3.19 5.30 CrackU-M 97.24 98.52 3.33 6.18 CrackU-L 97.79 98.81 3.98 9.76

[0135] Table 2 Comparison of effectiveness of different activation function combinations

[0136] ReLU GELU H-Swish mi IoU(%) mi Dice(%) √ 94.62 97.07 √ 96.24 97.96 √ 96.54 98.15 √ √ 96.93 98.36

[0137] Table 3 Performance comparison of different normalization methods

[0138] Standardized methods mi IoU(%) mi Dice(%) #Param.(M) MACs(G) BN 90.02 94.54 3.19 5.38 LN_PyTorch 84.91 91.66 82.60 5.30 LN_He 95.29 97.47 3.19 5.30 LN_VT 96.93 98.36 3.19 5.30

[0139] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A distributed Internet of Things system for mass concrete construction quality monitoring, characterized in that: include: The edge end and the cloud end, the edge end includes an online crack monitoring camera, an edge intelligent gateway, a temperature sensor and a strain sensor, and the cloud end is equipped with a computing server; The online crack monitoring camera includes an image acquisition module, an image processing module and a data transmission module; the image acquisition module is used to capture images of the concrete structure, the image processing module has an embedded crack recognition network and a crack quantitative characterization unit, which are used to identify cracks in the image and calculate crack information, respectively, and the data transmission module outputs the crack information to a computing server in the cloud; The temperature sensor and the strain sensor are arranged on the concrete structure and are used to detect the temperature data and the strain data of the concrete structure respectively; The edge intelligent gateway transmits the temperature data and strain data to a computing server in the cloud; The computing server performs computational analysis on the crack information, temperature data, and strain data, and sends instructions to the online crack monitoring camera and the edge intelligent gateway.

2. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 1, characterized in that: The embedded crack identification network is based on U-Net. The normalization layer and the activation function are not attached to the convolution layer at the same time. Only one of the normalization layer and the activation function is attached to the same convolution layer.

3. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 1, characterized in that: The normalization layer adopts the following layer normalization method: In the formula, The calculation result of the representation layer standardization, x i represents the input feature map, i represents the index of the input feature map, and Respectively represent the mean and standard deviation of the data set currently fed into the normalization layer; The activation function adopts the following H-Swish function: In the formula, a is the input of the activation function, ReLU6 is the ReLU function after maximum suppression, 4. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 1, characterized in that: The specific process of calculating the crack information by the crack quantitative characterization unit is as follows: Perform threshold segmentation on the crack recognition results of the image to determine whether each pixel in the image is a crack; Cracks are divided according to the connectivity between pixels, and the total number of cracks is counted; Calculate the total area of ​​the cracks; Extract the crack skeleton and calculate the total length of the crack; Calculate the average width of the cracks; The total area, total length and average width of the cracks are mapped from the pixel space to the physical space to obtain the actual total area, total length and average width of the cracks.

5. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 4, characterized in that: The crack recognition result of the image is subjected to threshold segmentation to determine whether each pixel in the image is a crack, specifically: The binary classification threshold is selected as 0.5, and the crack pixel-level recognition results are binarized as follows: In the formula, Pr t represents the prediction result of the model for the tth image, T represents the total number of images in the predicted image time series set, and Pr t (i, j) = 1 means that the pixel at position (i, j) in the image is predicted to be a crack, Pr t (i, j)=0 indicates that the position (i, j) is predicted to be a non-crack, and H and W represent the spatial height dimension and the spatial width dimension, respectively.

6. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 5, characterized in that: The cracks are divided according to the connectivity between pixels, specifically: Starting from any crack pixel, all crack pixels connected to it are found according to the crack connectivity criterion, and the total connected domain obtained represents a crack; after visiting all crack pixels, all cracks in the image are divided; The connectivity criterion of the crack is: if one pixel is a crack pixel and there are crack pixels in the surrounding eight pixels, then these crack pixels are interconnected; in one crack, all pixels are interconnected.

7. A distributed Internet of Things system for monitoring the quality of mass concrete construction as claimed in claim 5, characterized in that: The total area of ​​all cracks is calculated as follows: In the formula, Represents the total crack area of ​​the tth image Represents the result of binarization, and s represents the area of ​​a pixel in the image.

8. A distributed Internet of Things system for monitoring the quality of mass concrete construction as claimed in claim 7, characterized in that: The method of extracting the skeleton of each crack and calculating the total length of the crack is specifically as follows: The single-pixel-width skeleton of the crack is extracted, and the total length of the crack is calculated by the following formula: In the formula, Sk represents the total length of cracks in the tth image. t represents the skeleton graph of the t-th image, and l represents the side length of a pixel in the image.

9. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 8, characterized in that: The formula for calculating the average width of the crack is as follows: In the formula, represents the average crack width of the t-th image.

10. A distributed Internet of Things system for mass concrete construction quality monitoring as claimed in claim 9, characterized in that: The total area, total length and average width of the cracks are mapped from the pixel space to the physical space to obtain the actual total area, total length and average width of the cracks, specifically: Calibrate the online crack monitoring camera and calculate the proportionality coefficient ψ between the actual physical size and the image pixel; The actual total area, total length and average width of the cracks are calculated by the following formula: In the formula, and They represent the actual total area, total length and average width of the cracks respectively.

Images