Abnormal sampling early warning method and system based on concrete specimen setting state recognition

Abstract

The present application discloses an abnormal sampling warning method and system based on the identification of the setting state of concrete test blocks, wherein the method comprises: performing a rough identification on the extracted test block entity to classify the surface of the concrete test block as dry or wet, and then performing a fine identification on the test block entity with a classification result of surface wetness to classify the concrete test block as set or unset, thereby obtaining a setting state prediction result; wherein the test block entity is extracted from a captured image of the concrete test block surface; and comparing the normal setting state corresponding to the stage at which the concrete test block surface image was captured with the setting state prediction result, and if the comparison result is inconsistent, issuing a sampling abnormality warning signal. The present application quickly judges the setting state of the concrete test block surface image captured during the sampling stage, and accurately detects whether the concrete test block has an abnormality during the sampling stage based on the judgment result, thereby improving detection efficiency.

Description

Technical Field

[0001] The present application belongs to the technical field of sampling and identifying concrete test blocks at construction sites, and specifically relates to an abnormal sampling early warning method and system based on identification of the setting state of concrete test blocks. Background Art

[0002] With the continuous development of the modern construction industry, the quality of building materials plays a vital role in the safety and durability of buildings, and the quality of concrete test blocks is directly related to the safety of the main structure of the construction project. Concrete test blocks are produced on-site by staff, and witnesses prepare witness records to record the concrete test block information during the sampling, sample preparation, labeling, sealing, inspection, and on-site testing processes. During the sampling stage, in order to prevent the replacement of concrete test blocks, pre-made test blocks that have been treated with water or mortar are used as witness sampling test blocks to impersonate the test blocks that should be produced on-site. On-site personnel need to determine whether the concrete test blocks are in an unset state to determine whether there is any possibility of production abnormalities.

[0003] Existing anomaly detection methods primarily involve photographing the surface of concrete specimens during the sampling phase and encoding the image into a QR code. The sampler then uses the QR code to retrieve pre-entered sample information to determine whether the specimen is unset. This method of specimen status detection can determine the concrete setting state without inspecting the interior of the specimen and offers low photography costs. However, it relies heavily on manual labor, resulting in low efficiency and requiring high levels of professional expertise from witnesses. It is also highly subjective, lacks standardized standards, and is uncontrollable. Summary of the Invention

[0004] This application proposes an abnormal sampling early warning method and system based on the identification of the setting state of concrete test blocks. The setting state is quickly judged on the surface image of the concrete test block taken during the sampling stage. The judgment results are used to accurately detect whether the concrete test block has abnormalities during the sampling stage, thereby improving the detection efficiency.

[0005] A first aspect of the present application provides an abnormal sampling early warning method based on identification of the setting state of a concrete test block, the method comprising:

[0006] performing coarse recognition on the extracted test block entities to classify the surface of the concrete test block as dry or wet, and then performing fine recognition on the test block entities classified as wet to classify the concrete test block as set or unset, thereby obtaining a setting state prediction result; wherein the test block entities are extracted from the captured surface image of the concrete test block;

[0007] The normal setting state corresponding to the stage when the surface image of the concrete test block is taken is compared with the setting state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued.

[0008] After the concrete test block is completed, the above scheme photographs the surface of the test block to obtain a surface image. A series of preprocessing steps are then performed on the surface image to extract a complete, easily identifiable test block entity. Both coarse and fine identification are performed on the test block entity. In coarse identification, the image is first used to determine whether the test block surface is dry or wet. If the surface is wet, a more detailed fine identification is required to determine whether the interior of the test block is solidified or unsolidified. This allows for more accurate and efficient testing of the concrete test block's quality. This eliminates the need for inserting instruments into the test block for testing and eliminates the need for excessive reliance on manual judgment, improving the efficiency of project quality testing while reducing testing costs. The resulting solidification state prediction is then compared with the normal solidification state corresponding to the stage at which the concrete test block surface image was taken. If the concrete test block's solidification state does not match the expected normal state, the test block is considered likely to be replaced, and a sampling anomaly warning signal is issued, providing support for project quality assurance.

[0009] In a possible implementation method of the first aspect, the extracted test block entities are coarsely identified to classify whether the surface of the concrete test block is dry or wet. Then, the test block entities classified as having wet surface are finely identified to classify whether the concrete test block is set or unset, thereby obtaining a setting state prediction result, specifically:

[0010] Extracting features of the test block entity through a preset backbone network of a coagulation state recognition model to obtain a test block feature map;

[0011] Performing rough recognition on the test block characteristic map by using the rough recognition classifier of the coagulation state recognition model, and classifying the test block entity as a surface dry test block or a surface wet test block;

[0012] The test block feature map classified as a surface wet test block is precisely identified by using a transformer-based feature extraction network and a precise identification classifier of the coagulation state identification model, and the test block entity classified as a surface wet test block is classified as a coagulated water-sprayed test block, a coagulated mortared test block, or an uncoagulated test block;

[0013] According to the classification results of the coarse recognition and the classification results of the fine recognition, a prediction result of the setting state of the surface image of the concrete test block is obtained.

[0014] In a possible implementation method of the first aspect, feature extraction is performed on the test block entity through a preset backbone network of a coagulation state recognition model to obtain a test block feature map, specifically:

[0015] Downsampling the test block entity through the convolution module and the inverse residual module of the backbone network, and then extracting features from the downsampling results by superimposing several hybrid space interaction modules of the backbone network to obtain a test block feature map;

[0016] The hybrid space interaction module can be designed through actual feature extraction tasks.

[0017] The above scheme extracts features from the downsampling results by superimposing multiple hybrid space interaction modules designed based on actual feature extraction tasks. It can better capture the complex relationship between features, enhance the ability to extract the entity features of the test block in the image, and obtain a test block feature map with richer representation.

[0018] In a possible implementation method of the first aspect, the coarse recognition classifier of the condensation state recognition model performs coarse recognition on the test block characteristic map and classifies the test block entity as a surface dry test block or a surface wet test block, specifically:

[0019] The test block entity corresponding to the test block characteristic map is classified as a surface dry test block or a surface wet test block, and the test block characteristic map corresponding to the surface wet test block is input into the feature extraction network and the fine recognition classifier for fine recognition.

[0020] The above scheme uses a coarse recognition classifier to preliminarily classify the test block entities into surface wet test blocks and surface dry test blocks. Because the wetness and dryness of test blocks are relatively easy to distinguish, the coarse recognition method first classifies the test block entities corresponding to the surface dry test blocks with obvious and easily distinguishable features and the test block entities corresponding to the surface wet test blocks using the test block feature map, quickly determining which concrete test blocks are dry and may pose an abnormality risk.

[0021] In a possible implementation method of the first aspect, the test block feature map classified as a surface wet test block is precisely identified by using a transformer-based feature extraction network and a precise identification classifier of the coagulation state identification model, and the test block entity classified as a surface wet test block is classified as a coagulated water-sprayed test block, a coagulated mortared test block, or an uncoagulated test block, specifically:

[0022] Inputting the test block feature map classified as a surface wetness test block into a convolution module to adjust the number of image channels, and then inputting it into a feature extraction network based on a transformer structure to extract local features and global features to obtain a second feature map;

[0023] splicing the test block feature map, the feature map output by the convolution module, and the second feature map to obtain a feature fusion map corresponding to the surface wetness test block;

[0024] The feature fusion graph corresponding to the surface wetness test block is precisely identified by the precise identification classifier, and the test block entity classified as the surface wetness test block is classified as a condensed water spray test block, a condensed mortar test block or an uncondensed test block.

[0025] Compared to coarse recognition in the above scheme, fine recognition requires higher feature capture requirements. Therefore, a transformer-based feature extraction network and a fine recognition classifier are first used to more finely extract the features of the test block entity. The test block feature map, the feature map output by the convolution module, and the second feature map are then spliced ​​together. Through the interaction and fusion of feature information, a high-precision judgment is made as to whether the test block is internally solidified. Because the solidified and unsolidified states have overlapping feature information, it is difficult to judge through surface inspection. Therefore, a fusion of multiple feature maps is used here to integrate all comprehensive information for judgment. By combining local features with overall features, the unsolidified concrete test blocks can be more accurately identified.

[0026] In a possible implementation method of the first aspect, a prediction result of the setting state of the concrete test block surface image is obtained based on the classification result of the coarse recognition and the classification result of the fine recognition, specifically:

[0027] Classifying the concrete test block corresponding to the concrete test block surface image as a surface dry test block or a surface wet test block according to the classification result of the rough recognition;

[0028] According to the classification results of the precise identification, the concrete test blocks belonging to the surface wet test blocks are classified into solidified water-sprinkled test blocks, solidified mortared test blocks and unsolidified test blocks.

[0029] The above scheme first roughly divides the test blocks into dry and wet types based on the surface state of the test blocks through the coarse recognition results; then, through the fine recognition results, the test blocks with wet surfaces are further subdivided into internally condensed and uncondensed test blocks, accurately identifying the truly uncondensed test blocks.

[0030] In a possible implementation method of the first aspect, a comparison is performed between the correct setting state corresponding to the stage at which the surface image of the concrete test block is captured and the setting state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued, specifically:

[0031] If the stage is the sampling stage, the normal coagulation state is set to an uncoagulated test block;

[0032] If the setting state prediction result is an unset test block, the comparison results are consistent, and it is determined that the concrete test block corresponding to the setting state prediction result has no abnormality during the sampling stage;

[0033] If the setting state prediction result is a set test block, it means that the concrete test block corresponding to the setting state prediction result has an abnormality during the sampling stage, and a sampling abnormality warning signal is issued.

[0034] The above scheme first determines the characteristics of the concrete test block corresponding to the current testing stage, then compares the predicted setting state with the normal setting state corresponding to the stage when the concrete test block surface image was captured. If the comparison results are consistent, the test block has not been replaced and there is no risk of manufacturing abnormalities. If the comparison results are inconsistent, it indicates that a pre-made test block may have been used as a witness sample without treatment, watering, or grouting, to impersonate the test block that should have been produced on-site. In this case, a test block abnormality warning signal will be issued to warn.

[0035] In a possible implementation method of the first aspect, the surface dry test blocks, the condensed water spray test blocks and the condensed mortar test blocks are all classified as condensed test blocks.

[0036] In a possible implementation method of the first aspect, the test block entity is extracted from a captured surface image of the concrete test block, specifically as follows:

[0037] photographing the surface of the current concrete test block to obtain a plurality of images of the surface of the concrete test block;

[0038] Preprocessing the concrete test block surface image, and then inputting the preprocessed concrete test block surface image into an instance segmentation model for image segmentation to obtain a test block entity mask for each concrete test block surface image;

[0039] The test block entity is extracted from the concrete test block surface image according to the position of the test block entity mask in the concrete test block surface image.

[0040] A second aspect of the present application provides an abnormal sampling warning system based on concrete test block setting state identification, the system comprising: a test block state prediction module and a state comparison module;

[0041] The test block state prediction module is configured to perform coarse recognition on the extracted test block entities to classify the surface of the concrete test block as dry or wet, and then perform fine recognition on the test block entities classified as wet to classify the concrete test block as coagulated or uncoagulated, thereby obtaining a coagulation state prediction result; wherein the test block entities are extracted from the captured surface image of the concrete test block;

[0042] The state comparison module is used to compare the correct coagulation state corresponding to the stage when the surface image of the concrete test block is taken with the coagulation state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the implementation. Obviously, the drawings described below are only some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0044] Figure 1 This is a specific flow chart of an abnormal sampling early warning method based on concrete test block setting state identification provided by a certain embodiment of the present application;

[0045] Figure 2 This is an image cropping and comparison diagram of an abnormal sampling early warning method based on concrete test block setting state recognition provided by a certain embodiment of the present application;

[0046] Figure 3 This is a dry test block surface diagram of an abnormal sampling early warning method based on concrete test block setting state identification provided by a certain embodiment of the present application;

[0047] Figure 4 This is a graph of set and unset test blocks of an abnormal sampling early warning method based on identification of the setting state of concrete test blocks provided in one embodiment of the present application;

[0048] Figure 5 This is a concrete test block state recognition model diagram of an abnormal sampling early warning method based on concrete test block setting state recognition provided by a certain embodiment of the present application;

[0049] Figure 6 This is a structural diagram of the HSI Block, a basic module of an abnormal sampling early warning method based on concrete block setting state identification provided by an embodiment of the present application;

[0050] Figure 7 This is a structural diagram of an output head module of an abnormal sampling early warning method based on concrete test block setting state recognition provided by an embodiment of the present application;

[0051] Figure 8 This is a structural diagram of an abnormal sampling early warning system based on concrete test block setting state identification provided by a certain embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0053] It should be understood that the step numbers used herein are only for convenience of description and are not intended to limit the order in which the steps are to be executed.

[0054] First embodiment

[0055] In the field of engineering quality inspection, external markings, anti-counterfeiting patterns, etc. are often used to compare the test block information during witness sampling and sample verification to determine whether the test block has been replaced. During the sampling stage, there may be cases where pre-made test blocks are not processed, sprinkled with water, or slurried as test blocks for witness sampling, pretending that they should be test blocks made on site. This behavior of passing off inferior products as good ones will affect the authenticity of the test results and further affect the quality of the construction project. Therefore, in order to improve the inspection efficiency and accuracy of the witness sampling stage, the embodiment of the present application uses image recognition to quickly and accurately determine whether the concrete test blocks are in an unset state during the witness sampling stage, identify the test blocks in a set state, and issue an alarm signal in a timely manner.

[0056] like Figure 1 As shown, Figure 1 A specific flow chart of an abnormal sampling warning method based on concrete test block setting state identification is provided for a certain embodiment of the present application. The abnormal sampling warning method based on concrete test block setting state identification of this embodiment includes steps S1 to S3, which are detailed as follows:

[0057] Step S1: performing coarse identification on the extracted test block entity to classify the surface of the concrete test block as dry or wet, and then performing fine identification on the test block entity whose surface is classified as wet to classify the concrete test block as solidified or unsolidified, thereby obtaining a solidification state prediction result; wherein the test block entity is extracted from a captured image of the concrete test block surface.

[0058] According to relevant regulations, after the production of concrete test blocks is completed, witnesses are required to record the sampling, sample preparation, labeling, sealing, inspection, and on-site testing to ensure that the concrete test blocks are not replaced or replaced with inferior products, which would affect the quality of the building. During the sampling and sample inspection stages, witnesses will take photos of the concrete test blocks for record keeping, and use the photos to determine whether the test blocks are normal at the current stage. Because photos can be taken using only a mobile phone, it is a low-cost way to verify abnormalities in the test blocks, which can promote the development of the construction industry. Specifically, during the sampling stage, the concrete test blocks should be in an unset state; during the sample inspection stage, the concrete test blocks should be in a set state.

[0059] After vibrating and compacting concrete, it becomes an unsolidified, fluid, hydrated substance with a certain degree of fluidity. Its surface is in a moist gel state. During the curing process, cement hydrates, and the gel-state cement binds bulk materials such as sand and gravel together. During the curing process, it hardens, eventually forming a concrete block with a dry and solid surface. The embodiment of the present application learns the coagulation state characteristics of concrete test blocks through a deep learning model and summarizes the laws of different coagulation states. This allows for convenient, simple, direct, and efficient identification of the coagulation state of the test blocks. Moreover, compared to the prior art identification method of using a tool cone to penetrate a concrete test block, taking photos can ensure the integrity of the test block without causing damage to the test block.

[0060] If it is detected that the state of the test block in the image does not match the coagulation state corresponding to the stage, then the test block obviously has the possibility of abnormal production or delayed calculation age, and an early warning signal will be issued to prompt personnel to conduct further investigation.

[0061] This embodiment of the present application tests concrete test blocks during the sampling phase, which is when the concrete test blocks are in an unset state. During this phase, the surface of the completed concrete test blocks is photographed to obtain multiple images of the concrete test block surface. These concrete test block surface images are first preprocessed, primarily by resizing them to conform to the input format required by the instance segmentation model.

[0062] Optionally, in other embodiments, abnormal witness detection can be performed on concrete test blocks in other stages, such as the curing stage, the delivery stage, the inspection stage, the sample inspection stage, the compression resistance stage, etc.

[0063] Specifically, the preprocessing mainly involves scaling the surface image of the concrete specimen to the size specified by the instance segmentation model, converting it to RGB format, and standardizing the scaled image according to a certain mean and variance, and then converting the standardized image into Tensor format.

[0064] Optionally, in the embodiment of the present application, the mean of the image processing is set to 0.485, 0.456, and 0.406; the variance of the image processing is set to 0.229, 0.224, and 0.225.

[0065] The pre-processed concrete block surface images are input into the instance segmentation model for image segmentation, obtaining a block entity mask corresponding to each concrete block surface image. If the block entity mask does not border the actual edge of the concrete block surface image, it indicates that the block was captured completely and meets the requirements for subsequent anomaly detection. If the block entity mask borders one actual edge of the concrete block surface image, or borders two opposite edges of the concrete block surface image, it indicates that the block is a triple molded test block and meets the requirements for subsequent anomaly detection.

[0066] Optionally, the instance segmentation model includes but is not limited to Mask R-CNN, Solov1, Solov2, Yolo series, Yolact, etc.

[0067] Finally, the test block entity mask is cropped. Based on the position of the test block entity mask in the concrete test block surface image, the filled areas in the concrete test block surface image due to preprocessing operations and the background captured during shooting are removed, and the test block entity that fully displays the concrete test block surface is extracted.

[0068] In order to better demonstrate how the test block entity is extracted in the embodiment of the present application, Figure 2 Provides image cropping comparison diagram. As shown in the figure, the left side is the photographed concrete test block surface image, and the right side is the test block entity extracted based on the concrete test block surface image. Compared with the left image, the right image removes the redundant background around it, which can more completely show the surface of the concrete test block, and there will be no redundant background area to interfere with the subsequent coagulation state detection. In addition, Figure 2 The mosaic part in the middle of the concrete test block is the QR code identification of the concrete test block.

[0069] In an embodiment of the present application, by collecting and studying image data of the entire life cycle of concrete test blocks from their initial production and curing to sample delivery for inspection, these data are input into a deep learning model to observe the different states and characteristics of the test blocks from unset to initial set to set, so that the model can accurately distinguish the dry and wet properties of the concrete test block surface.

[0070] Because the dry and wet properties of the surface are relatively easy to distinguish, but whether the interior of the test block is condensed is more difficult to distinguish, the embodiment of the present application first divides the concrete test blocks into surface dry test blocks and surface wet test blocks based on the test block entity. Surface dry test blocks are test blocks in a condensed state, while surface wet test blocks can be in various states, including uncondensed test blocks, condensed watered test blocks, and condensed mortared test blocks. Therefore, further differentiation is required to identify uncondensed test blocks.

[0071] Based on the setting state, concrete specimens are divided into surface dry specimens, set watered specimens, set mortared specimens, and unset specimens. Among them, surface dry specimens, set watered specimens, and set mortared specimens are all classified as set specimens.

[0072] The main characteristics of surface dry test blocks are that the test block surface generally has cracks and voids, the edges are generally broken, there are cracks between the edges and the mold, there are generally sand or dry crushed concrete particles on the surface, it is generally off-white, and there may be writing or engraving. Figure 3 The surface image of the dry test block is provided. In the image, the surface of the concrete test block is non-reflective and has writing on it. There are sand and granular objects on the surface. There are cracks and holes on the edge of the test block, indicating that it is a surface dry test block. In addition, Figure 3 The mosaic part in the middle of the concrete test block is the QR code identification of the concrete test block.

[0073] The main characteristics of the condensed water test block are that the surface of the test block is generally dark gray and may have reflective properties. The other characteristics are similar to those of the surface dry test block.

[0074] The main characteristics of the solidified slurry test block are that there may be a layer of turbid gelatinous cement slurry on the surface of the test block, which is in the shape of a smooth spread, but it can be seen that there is a grayish-white solidified test block underneath the cement slurry; the original voids and cracks on the surface of the test block are filled with mud, and the area may appear darker in color; the cement slurry is generally reflective, and the edges are generally broken or there are cracks with the mold.

[0075] Both the condensed water-sprayed test blocks and the condensed mortared test blocks are condensed test blocks. The two have some overlapping characteristics. For example, they are dry but appear wet on the surface. The water / wet areas on the test block surface will reflect light. The difference between the two is:

[0076] (1) The water on the water-sprayed test block is clear and transparent, and does not affect the observation of the test block surface. Holes, cracks, and undulating wrinkles can be clearly seen; however, the cement slurry on the slurry test block is turbid, covering most of the test block surface, and almost no holes or cracks can be seen on the surface;

[0077] (2) The slurry of the test block is usually unevenly applied. Some areas have more slurry, while some areas have less or even no slurry. The color of the slurry area and the surface area of ​​the test block are generally quite different, and the color distribution is uneven.

[0078] (3) The slurry test block will generally have traces of the smoothing action, leaving traces or flat areas around the chip on the surface of the test block;

[0079] (4) The test block for screeding is prepared in advance, and the surface of the test block is flush with the mold. Therefore, the amount of cement slurry used for screeding is small, very thin, and generally does not contain high sand (otherwise it will protrude and not meet the sampling requirements). The appearance of the slurry is very different from that of the surface of the normal unset test block;

[0080] (5) The crack between the test block and the mold can be clearly seen in the water-sprayed test block, and the slurry-sprayed test block has a gully or height difference filled with mud;

[0081] (6) The mold of the slurry test block is usually smeared with slurry, which overflows onto the mold.

[0082] The above-mentioned distinguishing points can be used as distinguishing features between the solidified water-sprayed test blocks and the solidified mortared test blocks so that the solidification state recognition model can be learned and accurately distinguish between the unsolidified test blocks, the solidified water-sprayed test blocks, and the solidified mortared test blocks.

[0083] The main characteristic of an unset test block is that if there is a lot of water on the surface of the test block, it is generally because the test block has just been made, the water is turbid and gelatinous, there are generally no holes or cracks but there are bubbles, it is easy to produce reflections, and it is generally dark gray or bright yellow; if there is less water, it is generally because the test block is in the initial setting state, generally with less reflection, large holes but a small number, cracks but a small number, and there may be dry crushed concrete particles, etc., and it is generally dark gray or grayish white or bright yellow.

[0084] Figure 4 The graphs of the coagulated and uncoagulated test blocks are provided. The first row of the graph is the surface graph of the coagulated water-sprayed test block. It can be seen that the surface has a certain degree of reflection and the overall color is darker. The surface is relatively flat and clear. The second row is the surface graph of the coagulated slurry test block. It can be seen that the mud is unevenly distributed and some mud overflows onto the mold. The third row is the surface graph of the coagulated test block. It can be seen that there is a lot of water on the surface, which is turbid and gelatinous. The mud is also evenly distributed and easy to reflect light. In addition, Figure 4 The mosaic part in the middle of the concrete test block is the QR code identification of the concrete test block.

[0085] Based on the above description of characteristics, it can be seen that surface-dry test blocks have obvious characteristics, are relatively easy to identify, and are numerous, while surface-wet test blocks include a wider range of cases and are more difficult to identify. Therefore, to address this situation, the present embodiment of the application has designed a two-stage concrete test block setting state recognition strategy of coarse recognition + fine recognition. The test block entity is input into the setting state recognition model to first distinguish surface-dry test blocks from surface-wet test blocks through coarse recognition, and then distinguish unset test blocks from surface-wet test blocks through fine recognition.

[0086] The specific structure diagram of the condensation state recognition model provided in the embodiment of the present application is as follows Figure 5 As shown in the figure, Backone is the backbone network. The backbone network extracts the test block feature map from the input test block entity. The test block feature map is then used in the First Step to perform rough recognition to preliminarily classify the test block entity into surface dry test blocks and surface wet test blocks. The test block feature map is input into the Second Step for fine recognition to obtain the condensation state prediction result of each test block entity.

[0087] The backbone network's main function is to extract test block features. It consists of a convolutional block, an inverted residual block, a hybrid spatial interaction block (HSI block) designed based on task characteristics, and a head block. The backbone network extracts features from the test block entity. The feature extraction process is as follows:

[0088] 1. Downsampling operation is performed through the convolution module and the inverted residual module;

[0089] 2. By stacking multiple hybrid spatial interaction modules (HSI Block), features are fully extracted. The HSI Block uses 5×5 and 7×7 depthwise separable convolutions to implement medium- and long-distance spatial modeling, respectively, and then uses dot multiplication operations to achieve high-order spatial interaction, thereby better capturing the complex relationships between features and enhancing the ability to extract the entity features of the test blocks in the image. Residual splicing is then performed to superimpose new features on the original features to obtain a richer representation.

[0090] Among them, the structure of HSI Block is as follows Figure 6As shown, the output feature map of the previous module is expanded to 3 times the channel after 1×1 convolution, and then split into three parts. The first part is processed by 5×5 depth-separable convolution and sigmoid, and then directly multiplied with the second part; the result of the dot multiplication is processed by 1×1 convolution, and then the dot multiplication operation is performed with the third part after 7×7 depth-separable convolution and sigmoid processing; the dot multiplication result is spliced ​​with the output of the previous module after 1×1 convolution to obtain a feature map with 2 times the channel; finally, it is processed by 1×1 convolution and SE attention module to obtain the final output. In addition, Figure 6 CBA is a convolution layer-batch normalization layer-activation function structure, the full name is Convolution-BatchNorm-Activation, which is composed of a convolution layer, a batch normalization layer, and an activation function; DBA is a depth-separable convolution layer-batch normalization layer-activation function structure, the full name is Deconvolution-BatchNorm-Activation, which is composed of a depth-separable convolution layer, a batch normalization layer, and an activation function.

[0091] 3. Finally, the output result of the previous part is convolved with the output of the second inverted residual structure and then concatenated with the output after four HSI blocks, thereby integrating the low-dimensional basic and general feature information with the high-dimensional abstract and complex representation, improving the model performance and outputting the final test block feature map.

[0092] Next, the obtained test block feature map is used for the first stage of rough recognition through the Head Block, and the surface dry test blocks with obvious features and easy to distinguish are first classified.

[0093] Specifically, the coarse recognition is achieved by a coarse recognition classifier, which mainly includes a HeadBlock, which can perform simple classification of the test block entity through the test block feature map, and identify the test block feature map corresponding to the surface dry test block and the test block feature map corresponding to the surface wet test block. The structure of the Head Block of the coarse recognition classifier is as follows: Figure 7 shown.

[0094] Based on the previous description of the surface dry test block, the coarse recognition classifier can complete the classification by identifying whether the surface of the concrete test block is reflective, whether the test block is demoulded, whether there is a crack between the test block and the mold, etc.

[0095] According to the output results of the coarse recognition classifier Head Block, the test block feature maps representing the surface wetness of the test blocks are distinguished, and these feature maps are input into the second stage for fine recognition.

[0096] First, the feature map of the test block classified as a surface wetness test block is input into the convolution module to adjust the number of image channels. Then, the feature map output by the convolution module is input into a transformer-based feature extraction network to extract local and global features, thereby obtaining a second feature map. The second stage mainly consists of a convolution module, a transformer-based feature extraction network, and an output head (i.e., a precise recognition classifier). The convolution module can adjust the feature map channels, while the transformer module has long-range spatial modeling capabilities and high-order spatial interaction characteristics, which effectively capture the local and global features of the test block image, achieving accurate classification of the difficult-to-distinguish uncondensed and condensed watered and slurryed test blocks.

[0097] Specifically, the test block feature map output by the backbone network is first passed through the convolution module to change the number of channels so that the feature map meets the channel number requirements of the transformer module. Then the transformer module first expands the feature map output by the convolution module, regards the pixel points in the k×k area as a token, and changes the shape of the test block feature map from (B, C, H, W) to (B, C×k×k, H / k, W / k). Then, a 1×1 convolution is used to modify the channel to C, and then the dimensions are transformed and merged. At this time, the shape of the test block feature map is (B, H / k×W / k, C); the self-attention transformer structure is input for calculation to perform global information interaction and integration; then the test block feature map is folded, and after a dimensional transformation and 1×1 convolution opposite to the expansion operation, the shape of the test block feature map is changed from (B, H / k×W / k, C) to the original shape (B, C, H, W), and finally the constructed second feature map is output.

[0098] Then, the test block feature map output by the backbone network, the feature map output by the convolution module, and the second feature map output by the feature extraction network based on the transformer structure are spliced ​​and input into the convolution module for interaction and fusion of feature information. It not only includes the original information extracted by the backbone network, but also includes the long-distance spatial interaction information and local feature information of the test block features. The overall information is fully integrated for classification operations to obtain the feature fusion map corresponding to the surface wetness test block.

[0099] Finally, a precise recognition classifier is used to perform precise recognition on the feature fusion image, identifying unset test blocks, set watered test blocks, and set mortared test blocks. This precise recognition is achieved by considering the surface color of the concrete test blocks, the shape of the slurry on the test block surface, the number and size of holes on the test block surface, and the number of bubbles on the test block surface.

[0100] Therefore, fine recognition mainly uses the transformer structure to extract and integrate global feature information based on the CNN convolutional neural network of coarse recognition.

[0101] Step S2: comparing the correct setting state corresponding to the stage when the surface image of the concrete test block is taken with the setting state prediction result; if the comparison results are inconsistent, issuing a sampling abnormality warning signal.

[0102] In an embodiment of the present application, based on the coagulation state prediction results obtained from the coagulation state identification model, the concrete test blocks are divided into surface dry test blocks, coagulated watered test blocks, coagulated mortared test blocks and coagulated test blocks, among which the surface dry test blocks, coagulated watered test blocks and coagulated mortared test blocks are merged into coagulated test blocks.

[0103] For the sampling stage, the corresponding normal test block state should be the unset state. Therefore, by judging whether the set state prediction result is the unset state, it is determined whether the set state of the concrete test block in the current concrete test block surface image is normal. Specifically, there are two situations:

[0104] 1. When the concrete test block surface image taken during the sampling phase shows that the concrete test block is in a solidified state, it indicates that the pre-prepared test block may be used as the witness sampling test block without treatment, watering or slurrying, or the witness sampling may be delayed for at least one day after the on-site sampling. A test block abnormality warning signal should be issued;

[0105] 2. When the concrete test block is in an unset state in the surface image of the concrete test block taken during the sampling phase, it indicates that the state is normal during the sampling phase, and the setting state prediction result is returned.

[0106] The implementation of the embodiments of the present application has the following beneficial effects:

[0107] After the concrete test block is produced, the surface of the test block is photographed to obtain a concrete test block surface image. The concrete test block surface image is then segmented, and background and fill areas that are irrelevant to or interfere with the prediction are deleted from the image to extract a complete, easily identifiable test block entity. The test block entity is then subjected to coarse and fine recognition. In coarse recognition, the image is first used to determine whether the test block surface is dry or wet. If the test block surface is wet, a more detailed fine recognition is required to determine whether the interior of the test block has solidified or not. This allows for more accurate and efficient testing of the quality of the concrete test block. This eliminates the need to insert an instrument into the test block for testing and eliminates the need to rely heavily on human judgment, thereby improving the efficiency of engineering quality testing and reducing testing costs. The resulting solidification state prediction result is then compared with the normal solidification state corresponding to the stage at which the concrete test block surface image was taken. If the solidification state of the concrete test block does not match the normal state that should be present, it is considered that the test block may be replaced, and a sampling abnormality warning signal is issued to provide support for ensuring engineering quality.

[0108] Second embodiment

[0109] Furthermore, in order to implement the abnormal sampling warning system based on concrete test block setting state recognition corresponding to the above method embodiment to achieve the corresponding functions and technical effects, Figure 8 A structural diagram of an abnormal sampling warning system based on concrete block setting state identification is provided. For ease of illustration, only the parts relevant to this embodiment are shown. The abnormal sampling warning system based on concrete block setting state identification provided by the embodiment of the present application includes:

[0110] The test block state prediction module 201 is used to perform coarse identification on the extracted test block entities to classify the surface of the concrete test block as dry or wet, and then perform fine identification on the test block entities classified as wet to classify the concrete test block as set or unset, thereby obtaining a setting state prediction result; wherein the test block entities are extracted from the captured concrete test block surface image.

[0111] The state comparison module 202 is used to compare the normal setting state corresponding to the stage when the surface image of the concrete test block is captured with the setting state prediction result, and issue a sampling abnormality warning signal if the comparison results are inconsistent.

[0112] In some embodiments, the test block state prediction module 201 is specifically:

[0113] According to relevant regulations, after the production of concrete test blocks is completed, witnesses are required to record the sampling, sample preparation, labeling, sealing, inspection, and on-site testing to ensure that the concrete test blocks are not replaced or replaced with inferior products, which would affect the quality of the building. During the sampling and sample inspection stages, witnesses will take photos of the concrete test blocks for record keeping, and use the photos to determine whether the test blocks are normal at the current stage. Because photos can be taken using only a mobile phone, it is a low-cost way to verify abnormalities in the test blocks, which can promote the development of the construction industry. Specifically, during the sampling stage, the concrete test blocks should be in an unset state; during the sample inspection stage, the concrete test blocks should be in a set state.

[0114] After vibrating and compacting concrete, it becomes an unsolidified, fluid, hydrated substance with a certain degree of fluidity. Its surface is in a moist gel state. During the curing process, cement hydrates, and the gel-state cement binds bulk materials such as sand and gravel together. During the curing process, it hardens, eventually forming a concrete block with a dry and solid surface. The embodiment of the present application learns the coagulation state characteristics of concrete test blocks through a deep learning model and summarizes the laws of different coagulation states. This allows for convenient, simple, direct, and efficient identification of the coagulation state of the test blocks. Moreover, compared to the prior art identification method of using a tool cone to penetrate a concrete test block, taking photos can ensure the integrity of the test block without causing damage to the test block.

[0115] If it is detected that the state of the test block in the image does not match the coagulation state corresponding to the stage, then the test block obviously has the possibility of abnormal production or delayed calculation age, and an early warning signal will be issued to prompt personnel to conduct further investigation.

[0116] This embodiment of the present application tests concrete test blocks during the sampling phase, which is when the concrete test blocks are in an unset state. During this phase, the surface of the completed concrete test blocks is photographed to obtain multiple images of the concrete test block surface. These concrete test block surface images are first preprocessed, primarily by resizing them to conform to the input format required by the instance segmentation model.

[0117] Optionally, in other embodiments, abnormal witness detection can be performed on concrete test blocks in other stages, such as the curing stage, the delivery stage, the inspection stage, the sample inspection stage, the compression resistance stage, etc.

[0118] Specifically, the preprocessing mainly involves scaling the surface image of the concrete specimen to the size specified by the instance segmentation model, converting it to RGB format, and standardizing the scaled image according to a certain mean and variance, and then converting the standardized image into Tensor format.

[0119] Optionally, the mean of the image processing set in the embodiment of the present application is 0.485, 0.456, and 0.406; the variance of the image processing set is 0.229, 0.224, and 0.225.

[0120] The pre-processed concrete block surface images are input into the instance segmentation model for image segmentation, obtaining a block entity mask corresponding to each concrete block surface image. If the block entity mask does not border the actual edge of the concrete block surface image, it indicates that the block was captured completely and meets the requirements for subsequent anomaly detection. If the block entity mask borders one actual edge of the concrete block surface image, or borders two opposite edges of the concrete block surface image, it indicates that the block is a triple molded test block and meets the requirements for subsequent anomaly detection.

[0121] Optionally, the instance segmentation model includes but is not limited to Mask R-CNN, Solov1, Solov2, Yolo series, Yolact, etc.

[0122] Finally, the test block entity mask is cropped. Based on the position of the test block entity mask in the concrete test block surface image, the filled areas in the concrete test block surface image due to preprocessing operations and the background captured during shooting are removed, and the test block entity that fully displays the concrete test block surface is extracted.

[0123] By collecting and studying image data of concrete specimens from the time they are first made, cured, and finally sent for inspection, this data is input into a deep learning model to observe the different states and characteristics of the specimens from unset to initially set to set, so that the model can accurately distinguish the dry and wet properties of the concrete specimen surface.

[0124] Because the dry and wet properties of the surface are relatively easy to distinguish, but whether the interior of the test block is condensed is more difficult to distinguish, the embodiment of the present application first divides the concrete test blocks into surface dry test blocks and surface wet test blocks based on the test block entity. Surface dry test blocks are test blocks in a condensed state, while surface wet test blocks can be in various states, including uncondensed test blocks, condensed watered test blocks, and condensed mortared test blocks. Therefore, further differentiation is required to identify uncondensed test blocks.

[0125] Based on the setting state, concrete specimens are divided into surface dry specimens, set watered specimens, set mortared specimens, and unset specimens. Among them, surface dry specimens, set watered specimens, and set mortared specimens are all classified as set specimens.

[0126] The main characteristics of surface dry test blocks are that the test block surface generally has cracks and voids, the edges are generally broken, there are cracks between the edges and the mold, there are generally sand or dry crushed concrete particles on the surface, it is generally off-white, and there may be writing or engraving.

[0127] The main characteristics of the condensed water test block are that the surface of the test block is generally dark gray and may have reflective properties. The other characteristics are similar to those of the surface dry test block.

[0128] The main characteristics of the solidified slurry test block are that there may be a layer of turbid gelatinous cement slurry on the surface of the test block, which is in the shape of a smooth spread, but it can be seen that there is a grayish-white solidified test block underneath the cement slurry; the original voids and cracks on the surface of the test block are filled with mud, and the area may appear darker in color; the cement slurry is generally reflective, and the edges are generally broken or there are cracks with the mold.

[0129] Both the condensed water-sprayed test blocks and the condensed mortared test blocks are condensed test blocks. The two have some overlapping characteristics. For example, they are dry but appear wet on the surface. The water / wet areas on the test block surface will reflect light. The difference between the two is:

[0130] (1) The water on the water-sprayed test block is clear and transparent, and does not affect the observation of the test block surface. Holes, cracks, and undulating wrinkles can be clearly seen; however, the cement slurry on the slurry test block is turbid, covering most of the test block surface, and almost no holes or cracks can be seen on the surface;

[0131] (2) The slurry of the test block is usually unevenly applied. Some areas have more slurry, while some areas have less or even no slurry. The color of the slurry area and the surface area of ​​the test block are generally quite different, and the color distribution is uneven.

[0132] (3) The slurry test block will generally have traces of the smoothing action, leaving traces or flat areas around the chip on the surface of the test block;

[0133] (4) The test block for screeding is prepared in advance, and the surface of the test block is flush with the mold. Therefore, the amount of cement slurry used for screeding is small, very thin, and generally does not contain high sand (otherwise it will protrude and not meet the sampling requirements). The appearance of the slurry is very different from that of the surface of the normal unset test block;

[0134] (5) The crack between the test block and the mold can be clearly seen in the water-sprayed test block, and the slurry-sprayed test block has a gully or height difference filled with mud;

[0135] (6) The mold of the slurry test block is usually smeared with slurry, which overflows onto the mold.

[0136] The above-mentioned distinguishing points can be used as distinguishing features between the solidified water-sprayed test blocks and the solidified mortared test blocks so that the solidification state recognition model can be learned and accurately distinguish between the unsolidified test blocks, the solidified water-sprayed test blocks, and the solidified mortared test blocks.

[0137] The main characteristic of an unset test block is that if there is a lot of water on the surface of the test block, it is generally because the test block has just been made, the water is turbid and gelatinous, there are generally no holes or cracks but there are bubbles, it is easy to produce reflections, and it is generally dark gray or bright yellow; if there is less water, it is generally because the test block is in the initial setting state, generally with less reflection, large holes but a small number, cracks but a small number, and there may be dry crushed concrete particles, etc., and it is generally dark gray or grayish white or bright yellow.

[0138] Based on the above description of characteristics, it can be seen that surface-dry test blocks have obvious characteristics, are relatively easy to identify, and are numerous, while surface-wet test blocks include a wider range of cases and are more difficult to identify. Therefore, to address this situation, the present embodiment of the application has designed a two-stage concrete test block setting state recognition strategy of coarse recognition + fine recognition. The test block entity is input into the setting state recognition model to first distinguish surface-dry test blocks from surface-wet test blocks through coarse recognition, and then distinguish unset test blocks from surface-wet test blocks through fine recognition.

[0139] The condensation state recognition model provided in this embodiment comprises three main components: BackOne, First Step, and Second Step. BackOne is the backbone network. The input test block entity is used to extract a test block feature map through the backbone network. The test block feature map is then used in the First Step for coarse recognition to preliminarily classify the test block entity into surface dry test blocks and surface wet test blocks. The test block feature map is then input into the Second Step for fine recognition to obtain a predicted condensation state result for each test block entity.

[0140] The backbone network's main function is to extract block features. It consists of a convolutional block, an inverted residual block, a hybrid spatial interaction block (HSI block) designed based on task characteristics, and a head block. The feature extraction process is as follows:

[0141] 1. Downsampling operation is performed through the convolution module and the inverted residual module;

[0142] 2. By stacking multiple hybrid spatial interaction modules (HSI Block), features are fully extracted. The HSI Block uses 5×5 and 7×7 depthwise separable convolutions to implement medium- and long-distance spatial modeling, respectively, and then uses dot multiplication operations to achieve high-order spatial interaction, thereby better capturing the complex relationships between features and enhancing the ability to extract the entity features of the test blocks in the image. Residual splicing is then performed to superimpose new features on the original features to obtain a richer representation.

[0143] 3. Finally, the output result of the previous part is convolved with the output of the second inverted residual structure and then concatenated with the output after four HSI blocks, thereby integrating the low-dimensional basic and general feature information with the high-dimensional abstract and complex representation, improving the model performance and outputting the final test block feature map.

[0144] Next, the obtained test block feature map is used for the first stage of rough recognition through the Head Block, and the surface dry test blocks with obvious features and easy to distinguish are first classified.

[0145] Specifically, coarse recognition is achieved through a coarse recognition classifier, which mainly includes HeadBlock, which can simply classify the test block entity through the test block feature map, and identify the test block feature map corresponding to the surface dry test block and the test block feature map corresponding to the surface wet test block.

[0146] Based on the previous description of the surface dry test block, the coarse recognition classifier can complete the classification by identifying whether the surface of the concrete test block is reflective, whether the test block is demoulded, whether there is a crack between the test block and the mold, etc.

[0147] According to the output results of the coarse recognition classifier Head Block, the test block feature maps representing the surface wetness of the test blocks are distinguished, and these feature maps are input into the second stage for fine recognition.

[0148] First, the feature map of the test block classified as a surface wetness test block is input into the convolution module to adjust the number of image channels. Then, the feature map output by the convolution module is input into a transformer-based feature recognition network to extract local and global features, thereby obtaining a second feature map. The second stage mainly consists of a convolution module, a transformer-based feature extraction network, and an output head (i.e., a precise recognition classifier). The convolution module can adjust the feature map channels, while the transformer module has long-range spatial modeling capabilities and high-order spatial interaction characteristics, which effectively capture the local and global features of the test block image, achieving accurate classification of the difficult-to-distinguish uncondensed and condensed watered and slurryed test blocks.

[0149] Specifically, the test block feature map output by the backbone network is first passed through the convolution module to change the number of channels so that the feature map meets the channel number requirements of the transformer module. Then the transformer module first expands the feature map output by the convolution module, regards the pixel points in the k×k area as a token, and changes the shape of the test block feature map from (B, C, H, W) to (B, C×k×k, H / k, W / k). Then, a 1×1 convolution is used to modify the channel to C, and then the dimensions are transformed and merged. At this time, the shape of the test block feature map is (B, H / k×W / k, C); the self-attention transformer structure is input for calculation to perform global information interaction and integration; then the test block feature map is folded, and after a dimensional transformation and 1×1 convolution opposite to the expansion operation, the shape of the test block feature map is changed from (B, H / k×W / k, C) to the original shape (B, C, H, W), and finally the constructed second feature map is output.

[0150] Then, the test block feature map output by the backbone network, the feature map output by the convolution module, and the second feature map output by the feature extraction network based on the transformer structure are spliced ​​and input into the convolution module for interaction and fusion of feature information. It not only includes the original information extracted by the backbone network, but also includes the long-distance spatial interaction information and local feature information of the test block features. The overall information is fully integrated for classification operations to obtain the feature fusion map corresponding to the surface wetness test block.

[0151] Finally, a precise recognition classifier is used to perform precise recognition on the feature fusion image, identifying unset test blocks, set watered test blocks, and set mortared test blocks. This precise recognition is achieved by considering the surface color of the concrete test blocks, the shape of the slurry on the test block surface, the number and size of holes on the test block surface, and the number of bubbles on the test block surface.

[0152] In addition, fine recognition mainly uses the transformer structure to extract and integrate global feature information based on the CNN convolutional neural network of coarse recognition.

[0153] In some embodiments, the state comparison module 202 is specifically:

[0154] According to the setting state prediction results obtained from the setting state identification model, the concrete specimens are divided into surface dry specimens, set watered specimens, set mortared specimens and set specimens. Among them, the surface dry specimens, set watered specimens and set mortared specimens are combined into set specimens.

[0155] For the sampling stage, the corresponding normal test block state should be the unset state. Therefore, by judging whether the set state prediction result is the unset state, it is determined whether the set state of the concrete test block in the current concrete test block surface image is normal. Specifically, there are two situations:

[0156] 1. When the concrete test block surface image taken during the sampling phase shows that the concrete test block is in a solidified state, it indicates that the pre-prepared test block may be used as the witness sampling test block without treatment, watering or slurrying, or the witness sampling may be delayed for at least one day after the on-site sampling. A test block abnormality warning signal should be issued;

[0157] 2. When the concrete test block is in an unset state in the surface image of the concrete test block taken during the sampling phase, it indicates that the state is normal during the sampling phase, and the setting state prediction result is returned.

[0158] The implementation of the embodiments of the present application has the following beneficial effects:

[0159] After the concrete test block is produced, the surface of the test block is photographed to obtain a concrete test block surface image. The concrete test block surface image is then segmented, and background and fill areas that are irrelevant to or interfere with the prediction are deleted from the image to extract a complete, easily identifiable test block entity. The test block entity is then subjected to coarse and fine recognition. In coarse recognition, the image is first used to determine whether the test block surface is dry or wet. If the test block surface is wet, a more detailed fine recognition is required to determine whether the interior of the test block has solidified or not. This allows for more accurate and efficient testing of the quality of the concrete test block. This eliminates the need to insert an instrument into the test block for testing and eliminates the need to rely heavily on human judgment, thereby improving the efficiency of engineering quality testing and reducing testing costs. The resulting solidification state prediction result is then compared with the normal solidification state corresponding to the stage at which the concrete test block surface image was taken. If the solidification state of the concrete test block does not match the normal state that should be present, it is considered that the test block may be replaced, and a sampling abnormality warning signal is issued to provide support for ensuring engineering quality.

[0160] The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above description is merely a specific embodiment of this application and is not intended to limit the scope of protection of this application. In particular, it should be noted that for those skilled in the art, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. An abnormal sampling early warning method based on concrete block setting state recognition, characterized in that: include: performing coarse recognition on the extracted test block entities to classify the surface of the concrete test block as dry or wet, and then performing fine recognition on the test block entities classified as wet to classify the concrete test block as set or unset, thereby obtaining a setting state prediction result; wherein the test block entities are extracted from the captured surface image of the concrete test block; Before the coarse identification of the test block entity, the test block entity is subjected to feature extraction by a preset backbone network of a coagulation state recognition model to obtain a test block feature map. Specifically, the test block entity is downsampled by a convolution module and an inverse residual module of the backbone network, and then the downsampled results are subjected to feature extraction by superimposing a plurality of hybrid space interaction modules of the backbone network to obtain a test block feature map. Among them, the backbone network consists of a convolution module, an inverted residual module, a hybrid space interaction module and an output head module; the hybrid space interaction module expands the channel to 3 times after the downsampling result is subjected to 1×1 convolution, and then splits the processed downsampling result into three parts, wherein the first part is subjected to 5×5 depth-separable convolution and sigmoid processing, and then directly performs a dot multiplication operation with the second part, and the result after the dot multiplication is subjected to 1×1 convolution, and then the dot multiplication operation is performed with the third part after 7×7 depth-separable convolution and sigmoid processing. The result of the dot multiplication with the third part is spliced ​​with the downsampling result after 1×1 convolution to obtain a feature map of 2 times the channel, and finally the feature map of 2 times the channel is processed by 1×1 convolution and SE attention module and the result is output; The normal setting state corresponding to the stage when the surface image of the concrete test block is taken is compared with the setting state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued.

2. The abnormal sampling early warning method based on concrete test block setting state identification according to claim 1 is characterized in that: The extracted test block entities are roughly identified to classify whether the surface of the concrete test block is dry or wet, and then the test block entities classified as having wet surface are finely identified to classify whether the concrete test block is solidified or unsolidified, thereby obtaining a solidification state prediction result, specifically: Extracting features of the test block entity through a preset backbone network of a coagulation state recognition model to obtain a test block feature map; Performing rough recognition on the test block characteristic map by using the rough recognition classifier of the coagulation state recognition model, and classifying the test block entity as a surface dry test block or a surface wet test block; The test block feature map classified as a surface wet test block is precisely identified by using a transformer-based feature extraction network and a precise identification classifier of the coagulation state identification model, and the test block entity classified as a surface wet test block is classified as a coagulated water-sprayed test block, a coagulated mortared test block, or an uncoagulated test block; According to the classification results of the coarse recognition and the classification results of the fine recognition, a prediction result of the setting state of the surface image of the concrete test block is obtained.

3. The abnormal sampling early warning method based on concrete block setting state identification according to claim 2 is characterized in that: The coarse recognition classifier of the coagulation state recognition model performs coarse recognition on the test block characteristic map and classifies the test block entity into a surface dry test block or a surface wet test block, specifically: The test block entity corresponding to the test block characteristic map is classified as a surface dry test block or a surface wet test block, and the test block characteristic map corresponding to the surface wet test block is input into the feature extraction network and the fine recognition classifier for fine recognition.

4. The abnormal sampling early warning method based on concrete test block setting state identification according to claim 2 is characterized in that: The transformer-based feature extraction network and the precise identification classifier of the coagulation state identification model are used to precisely identify the test block feature map classified as a surface wet test block, and classify the test block entity classified as a coagulated water-sprayed test block, a coagulated slurry test block, or an uncoagulated test block, specifically: Inputting the test block feature map classified as a surface wetness test block into a convolution module to adjust the number of image channels, and then inputting it into a feature extraction network based on a transformer structure to extract local features and global features to obtain a second feature map; splicing the test block feature map, the feature map output by the convolution module, and the second feature map to obtain a feature fusion map corresponding to the surface wetness test block; The feature fusion graph corresponding to the surface wetness test block is precisely identified by the precise identification classifier, and the test block entity classified as the surface wetness test block is classified as a condensed water spray test block, a condensed mortar test block or an uncondensed test block.

5. The abnormal sampling early warning method based on concrete test block setting state identification according to claim 2 is characterized in that: The prediction result of the setting state of the concrete test block surface image is obtained based on the classification result of the coarse recognition and the classification result of the fine recognition, specifically: Classifying the concrete test block corresponding to the concrete test block surface image as a surface dry test block or a surface wet test block according to the classification result of the rough recognition; According to the classification results of the precise identification, the concrete test blocks belonging to the surface wet test blocks are classified into solidified water-sprinkled test blocks, solidified mortared test blocks and unsolidified test blocks.

6. The abnormal sampling early warning method based on concrete test block setting state identification according to claim 1 is characterized in that: The normal setting state corresponding to the stage when the surface image of the concrete test block is taken is compared with the setting state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued, specifically: If the stage is the sampling stage, the normal coagulation state is set to an uncoagulated test block; If the setting state prediction result is an unset test block, the comparison results are consistent, and it is determined that the concrete test block corresponding to the setting state prediction result has no abnormality during the sampling stage; If the setting state prediction result is a set test block, it means that the concrete test block corresponding to the setting state prediction result has an abnormality during the sampling stage, and a sampling abnormality warning signal is issued.

7. The abnormal sampling early warning method based on concrete test block setting state identification according to any one of claims 1 to 6, characterized in that: Surface dry test blocks, condensed water-sprayed test blocks and condensed mortared test blocks are all classified as condensed test blocks.

8. The abnormal sampling early warning method based on concrete test block setting state identification according to claim 1 is characterized in that: The test block entity is extracted from the surface image of the concrete test block, specifically: photographing the surface of the current concrete test block to obtain a plurality of images of the surface of the concrete test block; Preprocessing the concrete test block surface image, and then inputting the preprocessed concrete test block surface image into an instance segmentation model for image segmentation to obtain a test block entity mask for each concrete test block surface image; The test block entity is extracted from the concrete test block surface image according to the position of the test block entity mask in the concrete test block surface image.

9. An abnormal sampling early warning system based on concrete block setting state recognition, characterized in that: include: Test block state prediction module and state comparison module; The test block state prediction module is configured to perform coarse recognition on the extracted test block entities to classify the surface of the concrete test block as dry or wet, and then perform fine recognition on the test block entities classified as wet to classify the concrete test block as coagulated or uncoagulated, thereby obtaining a coagulation state prediction result; wherein the test block entities are extracted from the captured surface image of the concrete test block; Before the coarse identification of the test block entity, the test block entity is subjected to feature extraction by the backbone network of the preset coagulation state recognition model to obtain a test block feature map. Specifically, the test block entity is downsampled by the convolution module and the inverse residual module of the backbone network, and then the downsampling results are subjected to feature extraction by superimposing a plurality of hybrid space interaction modules of the backbone network to obtain the test block feature map. Among them, the backbone network consists of a convolution module, an inverted residual module, a hybrid space interaction module and an output head module; the hybrid space interaction module expands the channel to 3 times after the downsampling result is subjected to 1×1 convolution, and then splits the processed downsampling result into three parts, wherein the first part is subjected to 5×5 depth-separable convolution and sigmoid processing, and then directly performs a dot multiplication operation with the second part, and the result after the dot multiplication is subjected to 1×1 convolution, and then the dot multiplication operation is performed with the third part after 7×7 depth-separable convolution and sigmoid processing. The result of the dot multiplication with the third part is spliced ​​with the downsampling result after 1×1 convolution to obtain a feature map of 2 times the channel, and finally the feature map of 2 times the channel is processed by 1×1 convolution and SE attention module and the result is output; The state comparison module is used to compare the normal coagulation state corresponding to the stage when the surface image of the concrete test block is taken with the coagulation state prediction result. If the comparison result is inconsistent, a sampling abnormality warning signal is issued.

Images