A concrete curing intelligent monitoring method based on visual enhancement and regret theory

Abstract

This invention discloses an intelligent monitoring method for concrete curing based on multi-scale visual enhancement and KAN-TabPFN dynamic decision-making, comprising the following steps: Step 1: Detecting and segmenting concrete regions in visible light images using an improved YOLO11 network; Step 2: Mapping the concrete regions to infrared images through registration, extracting the corresponding concrete surface temperature values; simultaneously collecting environmental information, constructing a unified environmental state vector, and forming a short-period sequence; Step 3: Modeling the nonlinear trend of the sequence using a KAN network, and outputting corresponding expert weights through TabPFN model inference; Step 4: Decoupling the expert weights and inputting them into a three-branch decision model based on regret theory, dynamically calculating probability thresholds in conjunction with the current state; Step 5: Real-time correction of regret values ​​and updating of inputs, forming closed-loop control logic. This invention significantly solves the problems of missed detection of micro-cracks and cold-start decision-making, achieving refined curing with controllable risks.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, complex system state modeling and dynamic threshold decision-making, and specifically to an intelligent monitoring method for concrete curing based on visual enhancement and regret theory. Background Technology

[0002] As a core structural material in modern infrastructure construction and building engineering, concrete's hydration heat control and humidity maintenance during the early curing stage directly determine the structural strength, impermeability, and durability after molding. Traditional concrete curing management mainly relies on manual periodic inspections, experience-based manual temperature sensing, and fixed-frequency watering. This extensive management method not only consumes a lot of manpower and water resources, but also often suffers from irreversible curing defects such as plastic shrinkage cracks, sandblasting, or insufficient strength when facing extreme and complex conditions such as high temperatures, strong winds, or low humidity due to delayed response or misjudgment. Although some projects have introduced temperature and humidity sensors for auxiliary monitoring in recent years, their judgment logic is mostly based on a single physical quantity threshold (such as setting a fixed humidity lower limit), lacking an intuitive perception of the overall physical state of the concrete surface and a deep analysis of the coupling effects of multi-dimensional environmental factors.

[0003] With the rapid evolution of industrial internet and artificial intelligence technologies, surface defect detection based on machine vision and intelligent decision-making based on time-series data have gradually become research hotspots. However, existing intelligent maintenance systems still face significant technical bottlenecks in practical applications: First, in visual inspection, existing convolutional neural networks (such as the traditional YOLO series) often use stride convolution when performing downsampling operations, resulting in the significant loss of micro-cracks and subtle texture features generated in the early stages of concrete in the feature map, and making it difficult to achieve pixel-level accurate segmentation and localization under complex lighting and dirty backgrounds. Second, in terms of state modeling and weight analysis, traditional state-space models (such as RNN and LSTM) are difficult to accurately fit the nonlinear physical function curves in the concrete hydration process, and mainstream weight learning algorithms (such as XGBoost) often rely on massive amounts of historical labeled data, making it difficult to adapt to real-world scenarios such as cold starts or small sample data at construction sites. Most importantly, in the decision-making and control stage, most existing automated systems adopt simple binary decision-making or static fuzzy logic, ignoring the risk and cost differences caused by decision errors. This decision-making model, which lacks dynamic threshold adjustment capabilities and risk balancing mechanisms, struggles to find the optimal balance between ensuring maintenance quality and conserving resources, and cannot meet the stringent requirements of high-standard projects for meticulous maintenance. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the present invention aims to provide an intelligent monitoring method for concrete curing based on visual enhancement and regret theory. This method utilizes an improved YOLO11 network to introduce non-destructive downsampling and dynamic upsampling mechanisms in binocular vision, solving the problem of detecting micro-cracks. It employs a Kolmogorov-Arnold network (KAN) to accurately fit the physical change trends of environmental parameters, and combines it with a TabPFN model to achieve weighted inference under small sample conditions. Finally, it dynamically calculates the risk threshold through a three-branch decision mechanism based on regret theory, achieving refined and intelligent curing behavior, and significantly reducing cracking risk and resource waste.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart monitoring method for concrete curing based on visual enhancement and regret theory includes the following steps: Step 1: Simultaneously acquire visible light and infrared images of the concrete area using a binocular camera; construct an improved YOLO11 target detection network; use the YOLO11 target detection network to detect and segment the concrete area in the visible light images; The SPD-Conv (Space-to-DepthConvolution) lossless downsampling module is introduced into the backbone network to replace the traditional stride convolution, preserving the features of tiny cracks; the EMA (EfficientMulti-ScaleAttention) attention module is embedded in the feature fusion layer to suppress environmental noise; and the DySample (DynamicUpsampling) dynamic upsampling module is introduced in front of the detection head to achieve pixel-level accurate segmentation and localization of micro-cracks on the concrete surface. Step 2: The concrete area segmented in Step 1 is registered to the infrared image through perspective transformation, and the surface temperature matrix is ​​extracted; environmental parameters (temperature, humidity, wind speed, and light intensity) are collected simultaneously to construct a multi-dimensional environmental state vector; a short-period environmental sequence is constructed using a sliding time window and input into the KAN (Kolmogorov-Arnold Networks) neural network; the KAN network uses a learnable B-spline activation function to fit the nonlinear variation trend of the environmental parameters and outputs a high-order environmental trend feature function; this solves the shortcomings of traditional MLP models in fitting physical laws. Step 3: Input the high-order environmental trend feature function and the current environmental state into the TabPFN (TabularPrior-DataFittedNetwork) general inference model; TabPFN is based on the pre-trained Transformer architecture, and can directly output the expert weights (i.e. feature importance) of various sensor parameters on maintenance needs for the current small sample data without the need for gradient descent training locally, realizing second-level cold start weight inference. Step 4: Construct a three-way decision model based on regret theory. Using the expert weights obtained in Step 3 and the environmental state as input, calculate the regret value for cracking caused by untimely maintenance and waste caused by over-maintenance. Dynamically calculate the probability threshold based on the Bayesian risk minimization principle. and It outputs three decision commands: immediate watering (Positive), delayed observation (Boundary), or no action (Negative); Step 5: The intelligent control terminal executes operations based on the three decision commands. For delayed observation commands, the system will automatically increase the sampling frequency and perform micro-spraying. The regret value matrix will be corrected in real time through a closed-loop feedback mechanism until the decision is transformed into a definite execution or stop signal, thereby achieving precise maintenance with minimal risk.

[0006] Step 1, specifically the image acquisition and detection segmentation operations, includes the following steps: Step 1.1: Simultaneous acquisition of dual-band images. Using a binocular camera, visible light images of the concrete surface are simultaneously acquired at the same time and spatial location. and infrared thermal imaging images This step ensures that the acquired images have a spatial correspondence, which facilitates accurate registration and temperature extraction in the future, providing a basic support for the fusion of concrete surface condition and temperature information.

[0007] Step 1.2: Use an improved YOLO11 network to detect and segment the concrete area, and then use the visible light image obtained in Step 1.1. Input YOLO11 network for testing: in, The coordinates of the bounding box of the detected concrete area are shown in the following form: in, , These represent the minimum coordinate values ​​of the concrete region along the x-axis and y-axis of the two-dimensional image plane, respectively. , These are the maximum coordinate values ​​along the corresponding axes, and these four coordinates together determine the rectangular boundary range of the concrete area; they are used for subsequent detailed analysis of local areas.

[0008] Step 1.3: Introduce the SPD-Conv (Space-to-DepthConvolution) lossless downsampling module into the YOLO11 backbone network. This addresses the issue of feature loss in traditional stride convolution downsampling of early-stage micro-cracks in concrete, replacing the stride convolutional layer in the first layer of the backbone network. SPD-Conv preserves all pixel information through a space-to-depth mapping. For the input feature map... The four sub-feature maps generated by the slicing operation are as follows: in Indicates the spatial dimensions (width or height) of the input feature map. This indicates an index slicing operation performed on the feature map (slicing stride of 2). These represent four sub-feature maps obtained by slicing from different starting positions.

[0009] Then, the data is concatenated along the channel dimension to obtain the output feature map. : in This indicates a channel splicing operation. This represents the feature map output after stitching; the stitched feature map retains all the subtle texture information of the original image, effectively preventing feature blurring at low resolution.

[0010] Step 1.4: In the feature extraction stage, an EMA (Efficient Multi-Scale Attention) module is introduced to suppress complex background noise at the construction site (such as mud spots and shadows); the EMA module groups the feature maps and processes them separately. and Convolutional branches extract features and utilize cross-space learning mechanisms to aggregate pixel-level dependencies: in, For the first Global aggregation features of channels The height of the feature map, The width of the feature map. For the first Each channel is in coordinates The pixel values ​​at the location; the final generated attention weight map is used to weight the original features; significantly enhancing the saliency of the crack features.

[0011] Step 1.5: Introduce the DySample (Dynamic Upsampling) module before the detection head to replace the traditional linear interpolation upsampling. Considering the elongated and curved shape of the cracks, DySample learns the offset and dynamically resamples the feature map based on the content, making the segmentation edges clearer. Given input features... and the learned offset Output feature map In position The value is: in, This indicates that the output feature map is in coordinates Pixel value at that location, and The network is located at the following positions. The learned horizontal and vertical sampling offsets, These are the resampling weighting coefficients. It is a relative index within the sampling window; this module eliminates the jagged effect caused by traditional upsampling and ensures pixel-level segmentation accuracy of concrete crack edges.

[0012] In step 2, the specific operations of infrared registration, environmental state vector and KAN trend modeling include the following steps: Step 2.1: Registration and mapping of infrared images of the concrete area, using the bounding box of the concrete area detected in Step 1.2. The detection area on the visible light image is mapped to the infrared image in pixel space using internal calibration parameters: in, This represents the temperature matrix corresponding to the extracted concrete surface. Represents an infrared thermal imaging image. Represent the bounding box of the concrete region; extract the temperature matrix corresponding to the concrete surface. , used for subsequent combination of environmental state vectors; Step 2.2: Real-time acquisition and normalization of multimodal environmental parameters; The following environmental parameters are collected through multiple types of sensors deployed in the construction area: air temperature relative humidity Wind speed Light intensity Perform standard normalization on each parameter: in, Represents the normalized i-th One environmental parameter value, Indicates the first One original environmental parameter value, This indicates the minimum value of the parameter within the historical sample period. This indicates the maximum value of the parameter within the historical sample period.

[0013] Step 2.3: Nonlinear modeling of environmental trends based on KAN (Kolmogorov-Arnold Networks); constructing short-period environmental sequences. Input the KAN network. The KAN network abandons the traditional approach of using fixed activation functions on nodes, instead using learnable activation functions on edges (weights); the inter-layer mapping formula of the KAN network is: in, This represents the network's output value (i.e., trend characteristics). Indicates the dimension of the input data. Indicates the input number of the first... One portion, This represents the outer summation aggregate function. Indicates connection to input components With hidden layer nodes Learnable univariate activation functions; Among them, univariate activation function B-spline basis function parameterization is used to accurately fit the physical laws governing changes in environmental parameters: in, Indicates the first One B-spline basis function, This represents the adaptively adjusted control coefficients during training; the higher-order trend features output in this step. It can accurately reflect the nonlinear derivative characteristics of temperature and humidity changes over time (such as the rate of water evaporation), making up for the shortcomings of traditional linear models in terms of fitting ability.

[0014] Step 3, the specific operations of TabPFN weight inference include the following steps: Step 3.1: Construct TabPFN inference input; convert the high-order trend features output from Step 2 into... With the current environment state vector Feature concatenation is performed to form a query sample to be inferred. ; Step 3.2: Perform expert weight inference using the TabPFN (TabularPrior-DataFittedNetwork) model; TabPFN is a general inference model based on the Transformer architecture and pre-trained on a large amount of prior datasets. It treats input samples as tabular data and directly calculates the contribution of each input feature to the decision result through a self-attention mechanism. in, Represents the attention weight matrix. Represents the query matrix. Represents the key matrix. Represents a value matrix. Indicates the dimension of the key vector (used to scale the dot product result). This represents the product of the query matrix and the transpose of the key matrix. This represents the normalized exponential function; Step 3.3: Extract the channel weights within the Transformer that show the highest interest in the input features, and generate an expert weight vector: in, They represent the 1st to the 1st. The importance weight value of each sensor parameter This represents the total dimension of environmental parameters; this step eliminates the need for model training on the construction site, thus solving the cold start problem caused by a lack of data in the early stages of the project and directly outputting high-confidence sensor weights.

[0015] Step 3.4: Establish a front-to-back coupling structure between the KAN and TabPFN models to realize the linkage mechanism from physical trend perception to feature weight inference; and decouple the TabPFN output before it is fed into the decision module to ensure the independence of the expert scoring vector and the real-time state parameters.

[0016] In step 4, the specific operation of the three-way decision-making based on regret theory includes the following steps: Step 4.1: Construct the decision state set and action set; define the state set. , representing those requiring maintenance and those not requiring maintenance, respectively; define the action set. These represent immediate watering, no action required, and delayed observation, respectively. Step 4.2: Construct the regret value matrix and calculate the dynamic threshold; introduce a regret function to quantify the psychological cost of decision-making errors. in, Indicates the regret value. This represents the difference in utility resulting from the decision. The risk aversion coefficient ( >0), The base of the natural logarithm; Based on the Bayesian risk minimization principle, we derive the dynamic probability threshold pair. ; Calculate the lower probability threshold for immediate execution. : in, This represents the lower threshold for determining the probability of immediately watering. This represents the regret loss value incurred from taking no action when maintenance is required. This represents the regret loss value incurred from taking immediate action to water the plant when it does not require maintenance. This represents the regret loss value resulting from taking delayed observation actions when maintenance is required. This represents the regret loss value (usually 0) resulting from taking immediate watering action when maintenance is required. Calculate the upper limit threshold of the probability that no operation is required. : in, This represents the upper threshold of the probability that no action is required. This represents the regret loss value (usually 0) incurred from taking no action when no maintenance is required. This represents the regret loss value incurred by taking immediate watering action when no maintenance is required; the formula achieves dynamic automatic adjustment of the threshold according to risk cost; Step 4.3: Generate three decision instructions. Based on the expert weights from Step 3, calculate the conditional probability that the current state requires maintenance. And output the final instruction according to the following logic: like The output is immediately positive. like The output requires no operation (Negative). like Output delayed observation (Boundary);

[0017] Step 4.4: The decision output is transmitted to the control module; at the same time, for the delayed observation results, the high-frequency micro-probe feedback mechanism is activated as a trigger signal to build a closed-loop structure of perception-decision-trial-re-decision, and to lay the decision basis for the feedback mechanism in step 5.

[0018] In step 5, the specific operation of closed-loop feedback and control includes the following steps: Step 5.1: Perform control operations based on three-way decision-making.

[0019] Upon receiving an immediate execution command, the intelligent controller immediately activates the spray pumps for full-coverage maintenance; When a command requiring no operation is received, the system remains in standby mode. Upon receiving a delayed observation command, the system enters a high-frequency micro-probe mode, automatically increasing the sensor sampling frequency by 2 times and controlling the nozzle to perform a short-term (e.g., 3 seconds) micro-spray test. Step 5.2: Dynamic Feedback and Regret Value Correction. In delayed observation mode, the system monitors the surface humidity response rate of the concrete in real time after micro-spraying. If the response speed is lower than the preset value (indicating that the surface is extremely dry and absorbs water quickly), the regret value for not watering will be dynamically increased. This reduces the threshold. This prompts the decision-making logic to shift from delayed observation to immediate execution, forming a closed-loop control logic of perception-decision-trial-re-decision.

[0020] The beneficial effects of this invention are: This invention introduces SPD-Conv lossless downsampling, EMA multi-scale attention, and DySample dynamic upsampling modules into the YOLO11 network. SPD-Conv converts spatial information into channel depth, effectively solving the problem of lost features of fine cracks in concrete caused by traditional stride convolution; the EMA module suppresses complex background noise such as mud stains and shadows at the construction site; and DySample enables pixel-level dynamic upsampling. This combination significantly improves the model's ability to capture early, minute defects and enhances the clarity and anti-interference capabilities of segmentation edges.

[0021] This invention continues and deepens the dual-band image fusion mechanism. By simultaneously acquiring visible light and infrared images through a binocular camera and performing precise spatial registration, the physical texture information and thermal distribution information of the concrete surface are fused and analyzed in a multi-dimensional manner. This enables real-time perception of the dual state of concrete appearance and temperature, making up for the lack of information dimensions in a single visible light image under low light or dusty / foggy environments.

[0022] This invention applies a three-way decision-making (3WD) mechanism based on regret theory to intelligent concrete curing. It introduces a regret function to quantify the risk difference between cracking caused by insufficient curing and waste caused by over-curing. A probability threshold is dynamically calculated using the Bayesian risk minimization principle, and a delayed observation buffer is set. This mechanism replaces traditional static thresholds or simple binary logic, achieving an optimal risk balance between ensuring curing quality and conserving resources, significantly enhancing the robustness of the decision-making process.

[0023] This invention constructs a cold-start intelligent architecture that combines KAN trend modeling with TabPFN weight inference. Utilizing the learnable B-spline activation function of the KAN (Kolmogorov-Arnold Networks), it accurately fits the nonlinear physical trends of temperature and humidity changes with minimal parameters. Combined with the TabPFN pre-trained model, expert weight inference is performed directly without historical data training, effectively solving the cold-start problem caused by the scarcity of data in the early stages of construction sites, and significantly improving the system's deployment efficiency and environmental adaptability.

[0024] This invention constructs a closed-loop feedback mechanism of perception-decision-exploration-re-decision. In particular, under the delayed observation state of the three-way decision, the system automatically triggers high-frequency sampling and micro-volume test spraying operations. By dynamically correcting the regret value matrix and decision threshold through real-time monitoring of surface humidity response, the system can ensure that it can respond quickly and accurately to sudden environmental changes and guarantee the dynamic stability of the maintenance process. Attached Figure Description

[0025] Figure 1 This is a simulation diagram of the algorithm of the present invention.

[0026] Figure 2 This is a flowchart of an intelligent monitoring algorithm for concrete curing based on visual enhancement and regret theory, according to the present invention.

[0027] Figure 3 This is a block diagram of the improved YOLO11 algorithm of this invention.

[0028] Figure 4 The diagram shows the machine learning modules for embedding KAN and TabPFN in the algorithm of this invention.

[0029] Figure 5 This is a diagram of the 3WD decision-making module of the algorithm of this invention based on regret theory. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings.

[0031] like Figures 2-5 As shown, a smart monitoring method for concrete curing based on visual enhancement and regret theory includes the following steps: Step 1: Simultaneously acquire visible light and infrared images of the concrete area using a binocular camera; detect and segment the concrete area in the visible light image using an improved YOLO11 network, introducing an SPD-Conv (Space-to-DepthConvolution) lossless downsampling module in the first layer of its backbone network to replace the traditional stride convolution, preserving the features of minute cracks; embed an EMA (EfficientMulti-ScaleAttention) attention module in the feature fusion layer to suppress environmental noise; introduce a DySample (DynamicUpsampling) dynamic upsampling module in front of the detection head to achieve pixel-level accurate segmentation and localization of fine cracks on the concrete surface; Step 2: The concrete area segmented in Step 1 is registered and mapped to the infrared image to extract the corresponding concrete surface temperature value; at the same time, various environmental information, including temperature, humidity, wind speed, and light, are collected, and the raw sensor data is normalized using a standardization method to construct a unified environmental state vector; the raw environmental data is organized into a short-period sequence using a time window method, which is directly used as the input sequence of the state space model for dynamic modeling of trend features. Step 3: Input the environmental state vector and temperature and humidity trend features obtained in Step 2 into the KAN (Kolmogorov-Arnold Networks) neural network to generate a multidimensional data representation matrix with high-order nonlinear physical laws. This representation matrix is ​​then fed into the TabPFN (TabularPrior-DataFittedNetwork) general inference model as input features. Without on-site training, it can automatically infer the importance of various sensor parameters and concrete state information in the decision of whether to water the concrete, output the corresponding expert weights, and perform decoupling processing to provide a basis for the next step of decision-making. In step 3, the further specific operations of the machine learning and state modeling module include the following steps: Step 3.1: Input the multidimensional environmental state vector and its short-period time series constructed in Step 2 into the KAN neural network. The KAN model abandons the traditional fixed activation function and uses a learnable B-spline function as the activation function on the edges to perform nonlinear physical modeling of the sequence, accurately capturing the derivative characteristics and nonlinear laws of various environmental parameters over time (such as the rate of water evaporation), and outputting a multidimensional data representation matrix (i.e., trend characteristics). ).

[0032] Step 3.2: Concatenate the data representation matrix output by KAN with the current environmental state and input it into the TabPFN model for inference. TabPFN is based on a pre-trained Transformer architecture, treating the input as a table query sample. It directly calculates the feature contribution using a self-attention mechanism, eliminating the need for gradient descent training based on a large amount of historical data, unlike XGBoost. This allows it to output the contribution weights of each sensor parameter to the watering decision. This process effectively solves the cold start problem in the early stages of the project.

[0033] Step 3.3: Normalize and structure the Attention weights output by TabPFN to form an expert scoring vector; this scoring vector represents the degree of influence of various environmental factors in the current decision-making scenario and provides expert input for the next step of the three-branch decision-making mechanism based on regret theory.

[0034] Step 3.4: Establish a front-to-back coupling structure between the KAN and TabPFN models to realize the linkage mechanism from physical trend perception to feature weight inference; and decouple the TabPFN output before it is fed into the decision module to ensure the independence of the expert scoring vector and the real-time state parameters, thereby enhancing the generalization ability of the overall system under environmental changes.

[0035] In step 3, the hardware and software platform configuration used in the experiment includes: Windows 10 operating system, 13th Gen Intel(R) Core(TM) i9-13900HX processor, NVIDIA GeForce RTX 4060 graphics processing unit, CUDA 11.1 as GPU computing platform, and PyTorch 1.9.0 to build a deep learning framework for implementing the training and inference of the improved YOLO11 segmentation model, and the deployment and verification of KAN and TabPFN models.

[0036] Step 4: Decouple the various sensor weights obtained in Step 3 and input them as expert weights into the three-way decision (3WD) model based on regret theory. The input is formed together with the environmental state vector at the current moment. The three-way decision mechanism introduces a regret function to quantify the risk difference between cracking caused by missed maintenance and waste caused by over-maintenance. Based on the Bayesian risk minimization principle, the probability threshold is dynamically calculated, and three decision instructions including Positive, Negative, and Boundary are output. In step 4, the further specific operations of the three-branch decision model based on regret theory include the following steps: Step 4.1: Input the decoupled expert weights obtained in Step 3 and the current environmental state vector into the three-branch decision model; the model calculates the conditional probability that the current state belongs to the maintenance-required category based on the weights. .

[0037] Step 4.2: Construct the regret value matrix and introduce the risk aversion coefficient. Calculate the regret loss value for taking different actions (watering, not watering, observing) under different states (e.g.) (This represents the regret value for not watering when watering should have been done); based on these loss values, a pair of probability thresholds are dynamically derived. (Immediate execution threshold) and (No threshold operation required) This enables the threshold to be automatically adjusted according to environmental risk costs.

[0038] Step 4.3: Compare conditional probabilities With dynamic threshold Based on the relationship, output a comprehensive judgment result: like The output is executed immediately; like No operation is required for the output; like Output delayed observations.

[0039] This mechanism allows the system to proactively explore uncertain areas, rather than making blind decisions.

[0040] Step 4.4: The decision output is transmitted to the control module; at the same time, for the delayed observation results, the high-frequency micro-probe feedback mechanism is activated as a trigger signal to build a closed-loop structure of perception-decision-trial-re-decision, and to lay the decision basis for the feedback mechanism in step 5.

[0041] Step 5: Based on the three decision results output in Step 4, the system executes the corresponding actions through the intelligent control module; if the output is immediate execution, the spraying device is automatically turned on; if the output is delayed observation, the system automatically increases the sensor sampling frequency and controls the nozzle to perform micro-spraying, monitors the response rate of concrete surface humidity in real time, dynamically corrects the regret value matrix and updates the decision input, forming a closed-loop control logic to improve the real-time performance and robustness of the decision. In summary, this invention provides an intelligent monitoring method for concrete curing based on multi-scale visual enhancement and KAN-TabPFN dynamic decision-making. The system simultaneously acquires visible light and infrared images using a binocular camera. By introducing SPD-Conv non-destructive downsampling, EMA multi-scale attention, and DySample dynamic upsampling modules into an improved YOLO11 network, it significantly enhances the pixel-level detection capability for micro-cracks in concrete and solves the problem of small target feature loss caused by traditional stride convolution. The system collects multi-source environmental parameters to construct a state vector, uses a KAN network to accurately fit the nonlinear physical trend of temperature and humidity changes, and combines the TabPFN model to directly infer expert weights under small sample conditions, realizing a cold-start intelligent architecture of trend modeling → weight inference. Subsequently, the weights are injected into a three-branch decision-making mechanism based on regret theory through decoupling. The decision threshold is dynamically adjusted by quantifying the risk of misjudgment, and a delayed observation buffer is set. The system, combined with a closed-loop feedback mechanism, automatically triggers micro-trials and parameter corrections in the boundary domain state, ensuring precise temperature and humidity control during concrete curing while minimizing risk. The overall system combines high precision, strong robustness, and water conservation, significantly improving concrete curing efficiency and making it suitable for various complex outdoor construction environments.

[0042] Figure 1 The simulation diagram shows the effect of the intelligent monitoring algorithm for concrete curing, and the dynamic response process of the algorithm within the cycle: Figure 1 (a) shows that the concrete surface is under typical temperature rise and water loss conditions, with the temperature gradually rising from 25°C to 35°C and the relative humidity decreasing from 65% to 40% due to evaporation. Figure 1 (b) demonstrates the core decision-making logic, based on the immediate execution threshold calculated using regret theory. As environmental risk increases, the probability dynamically decreases from 0.85 to around 0.75, reflecting the algorithm's tendency to avoid cracking risks. Meanwhile, the maintenance necessity probability curve falls between time steps 25 and 72. and Within the yellow boundary region; Figure 1 (c) The final decision signal was output. During time steps 38 to 65, the output signal was 1 (corresponding to delayed observation / micro-spraying), indicating that the system should perform high-frequency sampling and active detection. After time step 65, the probability curve broke through. When the threshold is reached, the output signal jumps to 2 (corresponding to immediate execution), and the system officially starts full-coverage maintenance, verifying the algorithm's dynamic adaptability in non-steady-state environments.

Claims

1. A method for intelligent monitoring of concrete curing based on visual enhancement and regret theory, characterized in that, Includes the following steps: Step 1: Simultaneously acquire visible light and infrared images of the concrete area using a binocular camera, and construct an improved YOLO11 target detection network; use the YOLO11 target detection network to detect and segment the concrete area in the visible light images; Step 2: Register the concrete area segmented in Step 1 to the infrared image through perspective transformation and extract the surface temperature matrix; synchronously collect environmental parameters and construct a multi-dimensional environmental state vector; construct a short-period environmental sequence using a sliding time window and input it into the KAN neural network; the KAN network uses a learnable B-spline activation function to fit the nonlinear change trend of environmental parameters and outputs a high-order environmental trend feature function. Step 3: Input the high-order environmental trend feature function and the current environmental state into the TabPFN general inference model; TabPFN directly outputs the expert weights of various sensor parameters on maintenance requirements for the current small sample data, realizing second-level cold start weight inference; Step 4: Construct a three-branch decision model based on regret theory, using the expert weights and environmental state obtained in Step 3 as inputs, to calculate the regret value for cracking caused by untimely maintenance and waste caused by over-maintenance; Dynamically calculate probability thresholds based on the Bayesian risk minimization principle. and It outputs three decision commands: water immediately, delay observation, or no action required; Step 5: The intelligent control terminal executes operations based on the three decision commands. For delayed observation commands, the system will automatically increase the sampling frequency and perform micro-spraying. The regret value matrix is ​​corrected in real time through a closed-loop feedback mechanism until the decision is transformed into a definite execution or stop signal, thereby achieving precise maintenance with minimal risk.

2. The intelligent monitoring method for concrete curing based on visual enhancement and regret theory according to claim 1, characterized in that, Step 1, specifically the image acquisition and detection segmentation operations, includes the following steps: Step 1.1: Simultaneous acquisition of dual-band images. Using a binocular camera, visible light images of the concrete surface are simultaneously acquired at the same time and spatial location. and infrared thermal imaging images ; Step 1.2: Use an improved YOLO11 network to detect and segment the concrete area, and then use the visible light image obtained in Step 1.

1. Input YOLO11 network for testing: in, The coordinates of the bounding box of the detected concrete area are shown in the following form: in, , These represent the minimum coordinate values ​​of the concrete region along the x-axis and y-axis of the two-dimensional image plane, respectively. , These are the maximum coordinate values ​​in the corresponding axis directions; Step 1.3: Introduce the SPD-Conv lossless downsampling module into the YOLO11 backbone network. The SPD-Conv lossless downsampling module retains all pixel information through spatial-to-depth mapping; for the input feature map... The four sub-feature maps generated by the slicing operation are as follows: in Indicates the spatial size of the input feature map. This represents the index slicing operation performed on the feature map. These represent four sub-feature maps obtained by slicing from different starting positions; Then, the data is concatenated along the channel dimension to obtain the output feature map. : in This indicates a channel splicing operation. This represents the feature map output after concatenation; Step 1.4: In the feature extraction stage, an EMA attention module is introduced to suppress complex background noise at the construction site; the EMA module groups the feature maps and processes them separately. and Convolutional branches extract features and utilize cross-space learning mechanisms to aggregate pixel-level dependencies: in, For the first Global aggregation features of channels The height of the feature map, The width of the feature map. For the first Each channel is in coordinates The pixel values ​​at each location; the final attention weight map is used to weight the original features; Step 1.5: Introduce the DySample dynamic upsampling module before the detection head. Addressing the elongated and curved shape of the crack, DySample learns the offset and dynamically resamples the feature map based on the content, making the segmentation edges clearer; given the input features... and the learned offset Output feature map In position The value is: in, This indicates that the output feature map is in coordinates Pixel value at that location, and The network is located at the following positions. The learned horizontal and vertical sampling offsets, These are the resampling weighting coefficients. This is the relative index within the sampling window.

3. The intelligent monitoring method for concrete curing based on visual enhancement and regret theory according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: Registration and mapping of infrared images of the concrete area, using the bounding box of the concrete area detected in Step 1.

2. The detection area on the visible light image is mapped to the infrared image in pixel space using internal calibration parameters: in, This represents the temperature matrix corresponding to the extracted concrete surface. Represents an infrared thermal imaging image. Represent the bounding box of the concrete region; extract the temperature matrix corresponding to the concrete surface. , used for subsequent combination of environmental state vectors; Step 2.2: Real-time acquisition and normalization of multimodal environmental parameters; The following environmental parameters are collected through multiple types of sensors deployed in the construction area: air temperature relative humidity Wind speed Light intensity Perform standard normalization on each parameter: in, Represents the normalized i-th One environmental parameter value, Indicates the first Original environmental parameter values, This indicates the minimum value of the parameter within the historical sample period. This indicates the maximum value of the parameter within the historical sample period; Step 2.3: Nonlinear modeling of environmental trends based on KAN; constructing short-period environmental sequences. Input a KAN network; the inter-layer mapping formula of the KAN network is: in, This represents the network's output value. Indicates the dimension of the input data. Indicates the input number of the first... One portion, This represents the outer summation aggregate function. Indicates connection to input components With hidden layer nodes Learnable univariate activation functions; Among them, univariate activation function B-spline basis function parameterization is used to accurately fit the physical laws governing changes in environmental parameters: in, Indicates the first One B-spline basis function, This represents the control coefficient that is adaptively adjusted during training.

4. The intelligent monitoring method for concrete curing based on visual enhancement and regret theory according to claim 3, characterized in that, In step 3, the TabPFN general inference model specifically operates in the following steps: Step 3.1: Analyze the higher-order trend features output in Step 2. With the current environment state vector Feature concatenation is performed to form a query sample to be inferred. ; Step 3.2: TabPFN is a general inference model based on the Transformer architecture and pre-trained on a large amount of prior datasets; it treats input samples as tabular data and directly calculates the contribution of each input feature to the decision result through a self-attention mechanism: in, Represents the attention weight matrix. Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension of the key vector. This represents the product of the query matrix and the transpose of the key matrix. This represents the normalized exponential function; Step 3.3: Extract the channel weights within the Transformer that show the highest interest in the input features, and generate an expert weight vector: in, They represent the 1st to the 1st. The importance weight value of each sensor parameter This represents the total dimensions of environmental parameters; Step 3.4: Establish a front-to-back coupling structure between the KAN and TabPFN models to realize the linkage mechanism from physical trend perception to feature weight inference; before inputting the TabPFN output into the decision module, decouple it to ensure the independence of the expert scoring vector and the real-time state parameters.

5. The intelligent monitoring method for concrete curing based on visual enhancement and regret theory according to claim 4, characterized in that, In step 4, the specific operation of the three-way decision-making based on regret theory includes the following steps: Step 4.1: Define the state set , representing those requiring maintenance and those not requiring maintenance, respectively; define the action set. These represent immediate watering, no action required, and delayed observation, respectively. Step 4.2: Introduce a regret function to quantify the psychological cost of decision-making errors: in, This represents the regret value. This represents the difference in utility resulting from the decision. The risk aversion coefficient ( >0), The base of the natural logarithm; Based on the Bayesian risk minimization principle, we derive the dynamic probability threshold pair. ; Calculate the lower probability threshold for immediate execution. : in, This represents the lower threshold for determining the probability of immediately watering. This represents the regret loss value incurred from taking no action when maintenance is required. This represents the regret loss value incurred from taking immediate action to water the plant when it does not require maintenance. This represents the regret loss value resulting from taking delayed observation actions when maintenance is required. This represents the regret loss value resulting from taking immediate watering action when maintenance is required; Calculate the upper limit threshold of the probability that no operation is required. : in, This represents the upper threshold of the probability that no action is required. This represents the regret loss value incurred from taking no action required when maintenance is not needed. This represents the regret loss value incurred by taking immediate watering action when no maintenance is required; the formula achieves dynamic automatic adjustment of the threshold according to risk cost; Step 4.3: Generate three decision instructions. Based on the expert weights from Step 3, calculate the conditional probability that the current state requires maintenance. And output the final instruction according to the following logic: like The output is executed immediately; like No operation is required for the output; like Output delayed observations; Step 4.4: Simultaneously, for the delayed observation results, the high-frequency micro probe feedback mechanism is activated as a trigger signal.

6. The intelligent monitoring method for concrete curing based on visual enhancement and regret theory according to claim 5, characterized in that, In step 5, the specific operation of closed-loop feedback and control includes the following steps: Step 5.1: Perform control operations based on three-way decision-making; Upon receiving an immediate execution command, the intelligent controller immediately activates the spray pumps for full-coverage maintenance; When a command requiring no operation is received, the system remains in standby mode. Upon receiving a delayed observation command, the system enters a high-frequency micro-probe mode, automatically increasing the sensor sampling frequency by 2 times and controlling the nozzle to perform a short-term micro-volume test spray. Step 5.2: In delayed observation mode, the system monitors the surface humidity response rate of the concrete in real time after the micro-spray test. ; If the response speed is lower than the preset value, the regret value for not watering will be dynamically increased. This reduces the threshold. This prompts the decision-making logic to shift from delayed observation to immediate execution.

Images