Robot concrete slump and expansion degree measurement method based on weak supervised learning

Abstract

The invention discloses a robot concrete slump and expansion measurement method based on weak supervised learning, which comprises the following steps: acquiring image, displacement and mechanical data of concrete, and integrating a semi-supervised k-nearest neighbor model and a semi-supervised support vector machine model in a Scikit-learning framework; the problem that in the prior art, the capacity of processing incomplete labeled data is insufficient is effectively solved, the model can be trained under the condition that only part of labeled data exists, through cross validation, the model can dynamically adjust the weight of the semi-supervised k-nearest neighbor model and the weight of the semi-supervised support vector machine model according to actual measurement data, and the accuracy of the semi-supervised k-nearest neighbor model and the weight of the semi-supervised support vector machine model is improved. Therefore, a high-precision prediction result is still kept under the condition of part of labeled data, the position and parameters are automatically and dynamically adjusted by means of a gradient optimization algorithm for the measurement position, the optimal measurement position and parameter configuration are determined according to the variance of multiple measurement results, and the most reliable slump and expansion numerical values are output. The automation degree of the measurement process is improved, and high precision and high repeatability of the measurement result are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of concrete performance monitoring. More specifically, the present application relates to a robot concrete slump and spread measurement method based on weakly supervised learning. BACKGROUND

[0002] Concrete slump and spread are important indicators for evaluating the workability of concrete. Traditionally, the measurement method of the indicators mainly relies on manual operation and experience judgment, usually through manual measurement using standard slump cylinder or spread instrument on the construction site. The measurement process includes placing concrete in the instrument, measuring the subsidence or spread diameter of the concrete after removing the mold, and thus obtaining the numerical value of the slump or spread. However, this traditional method has multiple limitations. First, manual operation is easily affected by human factors, such as the experience, technique of the operator and the environmental conditions at the time, which may cause large fluctuations in the measurement results, thereby affecting the accuracy and consistency of the data. Second, the traditional measurement method usually requires fully labeled data to ensure the reliability of the measurement results. However, in the actual construction environment, the acquisition of labeled data is often incomplete, which may be caused by the limited measurement samples, the complexity of the construction site or other uncontrollable factors. In such cases, the traditional algorithms and methods are difficult to effectively process incomplete or partially labeled data, resulting in inaccurate or large error measurement results.

[0003] In addition, with the continuous expansion of modern construction engineering scale, the efficiency problem of traditional manual measurement is increasingly prominent. Large construction sites usually require multiple and continuous measurements to monitor the performance of concrete. Manual measurement not only consumes time, but also requires a large amount of human resources, which not only increases the cost of construction, but also may cause errors or omissions in the measurement results due to fatigue or negligence of human operation. In the current environment where the requirements for construction quality and efficiency are increasingly high, this is particularly prominent.

[0004] Based on these backgrounds, existing technologies have made some explorations in terms of automatic measurement and data processing. For example, some researchers have tried to use robot technology combined with image processing algorithms to realize automatic measurement of concrete slump and spread. Although this method improves the level of automation to some extent, there are still some problems that cannot be ignored. First of all, the above method mostly relies on completely labeled training data set. However, in actual application, the problem of incomplete labeling of data has not been effectively solved, which leads to insufficient robustness and accuracy of the existing method in dealing with complex construction environment. Secondly, the existing robot measurement system often lacks the ability of dynamic adjustment of measurement position and parameters. In complex field environment, it cannot adaptively adjust according to real-time measurement results. When the reference system of the sample or the sensor is offset relative to the theoretical reference, the system error will be accumulated in the three dimensions of geometric projection, stress direction and sampling area, which will affect the accuracy and consistency of the measurement.

[0005] Specifically, as long as the reference system of the sample or the sensor is offset relative to the theoretical reference, the system error will be accumulated in the three dimensions of geometric projection, stress direction and sampling area. The definition of slump h depends on the vertical distance between the upper and lower ends of the cone center axis. If the sample O point is placed at a position deviating from the reference axis Δx, Δy, and the gravity direction forms an angle θ with the normal direction of the sensor, the measured value becomes h' = hcosθ, causing a relative error Secondly, the spread measurement usually extracts the edge diameter d through the overhead camera. Once the camera optical axis and the normal direction of the sample plane form an angle The perspective scaling factor formed in the field of view is κ = cosφ, so d' = κd, and the error ε d = 1-κ is linearly related to A small angle deviation is amplified in the large diameter scene. Thirdly, the surface of the concrete after collapse is not isotropic, and the maximum settlement area often deviates from the geometric center. If the sampling radius r s of the sensor does not overlap with the actual maximum displacement radius r m , the displacement sensor can only capture local information in the range of Δr = |r s -r m |, which leads to the truncation of the slump curve and underestimates the total settlement. Finally, the site base often has millimeter-level unevenness or vibration. Any micro-inclination relative to the reference plane will change the normal component F n of the force sensor = Fcosθ, which causes the force-displacement curve to drift, thereby affecting the feature calibration of the subsequent weakly supervised model. The uncertainty of the measurement position is transmitted through the four links of angle error, perspective distortion, sampling area offset and base disturbance, which leads to systematic deviation in the estimation of slump and spread.

[0006] In summary, the prior art has significant defects in the automated measurement and data processing of concrete slump and spread, mainly in the insufficient processing capability of incomplete labeled data, the limited automation of the measurement process, and the low measurement accuracy and consistency in complex environments, which leads to the difficulty of the prior art to meet the demand of modern construction engineering for efficient, accurate and automated measurement. SUMMARY

[0007] An object of the present application is to solve at least the above problems and provide at least the advantages to be described later.

[0008] Another object of the present application is to provide a robot concrete slump and spread measurement method based on weakly supervised learning to solve the technical problem of insufficient accuracy of the prior art in the automated measurement of concrete slump and spread.

[0009] In order to achieve these objects and other advantages according to the present application, a robot concrete slump and spread measurement method based on weakly supervised learning is provided, comprising the following steps:

[0010] S1: setting and initializing a robot system, configuring a connection sensor module for acquiring image data, displacement data, and mechanical data of the concrete to be measured;

[0011] S2: acquiring a concrete sample to be measured, placing the concrete sample to be measured at a designated position for measurement by the robot system, and automatically adjusting the concrete sample to a preset reference position by the robot system, with a deviation value not exceeding the set deviation value from the reference position;

[0012] S3: starting the sensor module to collect data and integrating to form a preliminary measurement data set of concrete slump and spread;

[0013] S4: after preprocessing the preliminary measurement data set, inputting it into an improved weakly supervised learning model for training, and outputting the prediction results of the concrete slump and spread;

[0014] S5: based on the prediction results of the improved weakly supervised learning model, calculating the initial measurement position, parameter configuration and adjusted measurement position, parameter configuration of the concrete sample each time by the robot system, obtaining the concrete slump and spread under different measurement positions and parameters, forming a sample data set, analyzing the sample data set, determining the optimal measurement position and parameter configuration according to the variance of the sample data set, and corresponding outputting the final numerical value of the concrete slump and spread and the determination result of whether it is qualified;

[0015] S6: visualizing and displaying the numerical value determined to be qualified in step S5.

[0016] Preferably, the sensor module comprises a vision camera for collecting image data, a displacement sensor for collecting displacement data, and a force sensor for collecting mechanical data, wherein the vision camera is arranged towards the concrete sample, the force sensor is installed below the measuring table, and the displacement sensor is a combination of a laser displacement meter and a vision structured light, the laser channel is used to monitor the settlement of the center point, and the structured light is used to reconstruct the surface grid and calculate the radial diffusion diameter.

[0017] Preferably, step S2 specifically comprises the following steps:

[0018] S21: After obtaining the concrete sample to be measured, the robot system automatically grabs the concrete sample through the mechanical arm, and moves the concrete sample to the designated position for measurement at a constant speed v c , wherein the designated position and the deviation value ∈ p are pre-set according to the measurement task, and the center of the concrete sample is aligned with the reference point of the measuring device;

[0019] S22: The actual position coordinates P s (x s ,y s ,z s ) of the concrete sample are measured by the displacement sensor, and the concrete sample is compared with the pre-set reference position coordinates P r (x r ,y r ,z r ):

[0020]

[0021] Wherein, ΔP represents the deviation value of the actual position of the concrete sample from the reference position;

[0022] S23: When ΔP>∈ p , the position of the concrete sample is fine-tuned by the robot system until ΔP≤∈ p ;

[0023] S24: After fine-tuning, the position of the concrete sample is fixed by the robot system.

[0024] Preferably, in the step S3:

[0025] The image data set of the concrete sample is obtained by the camera, and the image data set comprises a plurality of image frames at different angles:

[0026] I={I1,I2,…,I n}

[0027] Wherein, I i represents the i-th image frame, and n is the total number of frames.

[0028] Data set of displacement measurements of concrete samples at different time points using displacement sensors:

[0029] D = {d1, d2, ..., d} m}

[0030] Where, d j This represents the displacement value at the j-th time point, where m is the total number of measurement time points;

[0031] Mechanical data sets of concrete samples under loading conditions were obtained using force sensors.

[0032] F = {f1, f2, ..., f k}

[0033] Among them, f p This represents the force value at the p-th measurement point, where k is the total number of measurement points;

[0034] The image dataset I, displacement dataset D, and mechanical dataset F are integrated to form a preliminary measurement dataset:

[0035] M = {I, D, F}.

[0036] Preferably, step S4 specifically includes the following steps:

[0037] S41: Input the preprocessed measurement dataset into the semi-supervised learning model ensemble model based on the Scikit-learn framework for model training and validation, and calculate the prediction accuracy of the model. The Scikit-learn framework integrates a semi-supervised k-nearest neighbor model and a semi-supervised support vector machine model.

[0038] S42: Use the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework to initially train the measurement dataset and obtain their respective prediction results y. kNN and y SVM ;

[0039] S43: Integrate the outputs of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model through a voting process to generate the final prediction result y. final :

[0040] y final =α·y kNN +β·y SVM ;

[0041] Where α and β are the weight parameters of the corresponding model, and α+β=1;

[0042] S44: In the training process of the integrated model, the weight parameters a and β of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model are dynamically adjusted through cross-validation, so that the integrated model obtains the best prediction performance under the condition of partial labeled data;

[0043] S45: Finally output the prediction results of the integrated model on the concrete slump and spread, and classify and predict the unlabeled data to generate the final evaluation values of the concrete slump and spread corresponding to the data.

[0044] Preferably, S41 comprises the following steps:

[0045] S411: Divide the preprocessed measurement data set into a training data set M train and a test data set M test , the proportion of the division is determined according to the following formula:

[0046]

[0047] Wherein, r is the proportion of the training set, λ1 is the proportion control parameter, σ M is the standard deviation of the measurement data set, is the baseline standard deviation;

[0048] S412: Initialize the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework, and set the model parameters respectively. For the semi-supervised k-nearest neighbor model, the number of neighbors k is selected according to the following optimization:

[0049]

[0050] Wherein, η kNN is the trade-off parameter, is the feature vector of the i-th training sample, is the corresponding label, ||| represents the Euclidean distance, and l(·,·) represents the label loss function;

[0051] For the semi-supervised support vector machine model, the kernel function parameter γ is selected by the following adaptive optimization:

[0052]

[0053] Wherein, w SVM is the weight vector of the support vector machine, is the kernel function mapping of the input sample, μ SVM and v SVM are regularization parameters, is the model output using the kernel function parameter γ;

[0054] S413: Divide the training data set Mtrain The inputs are fed into a semi-supervised k-nearest neighbor model and a semi-supervised support vector machine model, respectively, for training. The trained semi-supervised k-nearest neighbor model and semi-supervised support vector machine model are obtained. The training process is optimized by minimizing the following comprehensive loss function:

[0055]

[0056] Where, λ kNN , λ SVM λ3 are the weight parameters of the loss function. Let be the loss functions corresponding to the k-nearest neighbor model and the support vector machine model, respectively;

[0057] S414: Using the test dataset M test The trained semi-supervised k-nearest neighbor model and semi-supervised support vector machine model were validated, and the prediction accuracy of the models was calculated:

[0058]

[0059] in, As an indicator function, when predicting labels With real labels ω is 1 if they match, otherwise it is 0. kNN and ω SVM These are the weight coefficients for the k-nearest neighbor model and the support vector machine model, respectively.

[0060] Preferably, the loss functions of the k-nearest neighbor model and the support vector machine model are designed by combining the prediction accuracy of concrete slump and spread, the feature distance between samples, and the time series error.

[0061] The loss function for the k-nearest neighbor model is as follows:

[0062]

[0063] in, This represents the concrete slump or spread predicted by the semi-supervised k-nearest neighbor model. λ is the actual measured value. d1 To control the weighting coefficient of the Euclidean distance between samples, λ d1 To control the weighting coefficients of time series errors, Let be the feature vector of the i-th sample. and These represent the model prediction and the actual measurement time, respectively.

[0064] The loss function for the support vector machine model is as follows:

[0065]

[0066] wherein, is the true concrete slump or spread label, is the predicted value generated using the support vector machine model, w SVM is the weight vector of the support vector machine, λ d2 is the regularization parameter that controls the model complexity, λ t2 is the weight coefficient that controls the temporal error, is the measured time of the model prediction, b is the bias term.

[0067] Preferably, S42 comprises the following steps:

[0068] S421 : input the feature data x i and the partially labeled target data into the semi-supervised k- nearest neighbor model in the Scikit-learn framework, and preliminarily train according to the k- nearest neighbor model loss function to obtain the prediction result y kNN of the k- nearest neighbor model:

[0069]

[0070] wherein, w j is the weight of the jth neighbor sample, y j is the predicted label of the sample, λ1 is the regularization parameter that adjusts the weight, x i and x j are the feature vectors of the sample to be measured and the neighbor sample, respectively;

[0071] S422: input the feature data x and the partially labeled target data into the semi-supervised support vector machine model in the Scikit-learn framework, and preliminarily train according to the support vector machine model loss function to obtain the prediction result y SVM of the support vector machine model:

[0072]

[0073] wherein, w SVM is the weight vector of the support vector machine, is the kernel function mapping of the input feature, b is the bias term, and λ2 is the regularization parameter.

[0074] Preferably, when labeling the data, a unified metadata structure (X, y, δ) is established for each sample at the data access layer, wherein X represents a three-modal feature tensor, y is a target vector, and δ is a labeling state indicator. When the test sample is measured by the artificial reference method, the slump h ref and the spread d ref are obtained.And when the verification error meets the specified ±2mm range, write y=(h ref ,d ref And set δ to 1. If the benchmark value is missing, out of tolerance, or still in the verification queue, then assign y to... Setting δ to 0 achieves physical separation between labeled and unlabeled flows.

[0075] Automatic construction according to δ and The former participates in the supervised part of the loss function of semi-supervised kNN / SVM. The latter uses consistency regularization and pseudo-labeling strategies to enter the unsupervised partial loss.

[0076] Preferably, S5 specifically includes the following steps:

[0077] A1: Output result y based on the improved weakly supervised learning model final The robot system calculates the initial measurement position P0(x0,y0,z0) of the concrete sample and its corresponding parameter configuration Θ0={θ1,θ2,…,θ n}, including the parameter configuration of the vision camera and laser used to acquire the position of the concrete sample, where P0 is the three-dimensional coordinate of the measurement position and Θ0 is the parameter set;

[0078] A2: The robot system automatically adjusts itself based on the initial measurement position P0 and parameter configuration Θ0, and calculates the adjusted measurement position P. adj (x adj ,y adj ,z adj ):

[0079] P adj =P0 + ΔP where

[0080] Where ΔP is the adjustment vector for the measurement position, and α3 is the adjustment step size. Represents the loss function for measurement location P. The gradient;

[0081] A3: Perform multiple measurements, and recalculate the adjusted measurement position P after each measurement. adj and the corresponding parameter configuration Θ adj To obtain sample datasets of concrete slump and spread under different measurement locations and parameters:

[0082]

[0083] A4: For sample dataset M meas The analysis was performed based on the variance Var(M) of the sample dataset. measdetermining the optimal measurement position P opt and parameter configuration Θ opt outputting the final concrete slump and spread values y opt and the determination result R opt :

[0084]

[0085]

[0086] wherein β3 is a weight coefficient, y min and y max are preset qualified determination ranges.

[0087] The present application at least includes the following beneficial effects:

[0088] (1) The present application develops an integrated system of semi-supervised learning models in the Scikit-learn framework, integrating semi-supervised k-nearest neighbor models and semi-supervised support vector machine models, effectively solving the problem of insufficient processing capacity for incomplete labeled data in the prior art. By introducing a semi-supervised learning algorithm, the model can be trained with only partial labeled data, reducing dependence on fully labeled data and improving the model's generalization ability. At the same time, by integrating multiple semi-supervised models and using voting or weighted averaging to generate the final prediction result, the model's prediction accuracy for slump and spread is further enhanced, significantly improving the accuracy and reliability of measurement results in complex construction environments.

[0089] (2) The present application adopts a cross-validation mechanism to dynamically adjust the weights of each base model, enabling the integrated model to adaptively adjust in different environments to achieve the best prediction results. Through cross-validation, the model can dynamically adjust the weights of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model based on actual measurement data, thereby maintaining high-precision prediction results under the condition of partial labeled data.

[0090] (3) The present application uses a robot system to automatically adjust the measurement position and parameters, and combines the output results of the weakly supervised learning model to perform multiple measurements and adjust the measurement strategy in real time. Through the gradient optimization algorithm for the measurement position, the robot can dynamically adjust the position and parameters after each measurement to obtain more accurate measurement data. In addition, the system determines the final optimal measurement position and parameter configuration based on the variance of multiple measurement results, thereby outputting the most reliable slump and spread values, improving the degree of automation in the measurement process, reducing the possibility of human intervention, and ensuring high precision and high repeatability of the measurement results.

[0091] Additional advantages, objects, and features of the application will be apparent from the following description, taken in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0092] Fig. 1 Flowchart of the present application;

[0093] Fig. 2 Schematic diagram of the integrated architecture of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework of the present application. DETAILED DESCRIPTION

[0094] The present application will be further described below in conjunction with the accompanying drawings, so that those skilled in the art can implement the present application according to the description and the drawings.

[0095] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials can be obtained commercially unless otherwise specified; in the description of the present application, the terms "lateral", "longitudinal", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0096] As shown in Figs. 1-2 The present application provides a robot concrete slump and spread measurement method based on weakly supervised learning, comprising the following steps:

[0097] S1: setting and initializing a robot system, configuring a connection sensor module for acquiring image data, displacement data, and mechanical data of the concrete to be measured.

[0098] Preferably, the sensor module includes a vision camera for collecting image data, a displacement sensor for collecting displacement data, and a force sensor for collecting mechanical data, wherein the vision camera is arranged towards the concrete sample, the force sensor is installed below the measurement table, and the displacement sensor is a combination of a laser displacement meter and a vision structured light, the laser channel is used to monitor the settlement of the center point, and the structured light is used to reconstruct the surface grid and calculate the radial diffusion diameter.

[0099] S2: obtaining a concrete sample to be measured, placing the concrete sample to be measured at a designated position for measurement by the robot system, and automatically adjusting the concrete sample to a preset reference position by the robot system, with a deviation value not exceeding the set deviation value from the reference position.

[0100] Step S2 specifically comprises the following steps:

[0101] S21: After obtaining the concrete sample to be measured, the robot system automatically grabs the concrete sample through the mechanical arm, and moves the concrete sample to the designated position for measurement at a constant speed v c , the designated position and the deviation value ∈ p are preset according to the measurement task, aligns the center of the concrete sample with the reference point of the measuring device;

[0102] S22: The actual position coordinates P s (x s ,y s ,z s ) of the concrete sample are measured by the displacement sensor, and the concrete sample is compared and calculated with the preset reference position coordinates P r (x r ,y r ,z r ):

[0103]

[0104] Where ΔP represents the deviation value of the actual position of the concrete sample from the reference position;

[0105] S23: When ΔP > ∈ p , the position of the concrete sample is fine-tuned by the robot system until ΔP ≤ ∈ p ;

[0106] S24: After fine-tuning, the position of the concrete sample is fixed by the robot system.

[0107] S3: Start the sensor module to collect data, control the sensor module to measure the slump and spread of the concrete sample, wherein the sensor module includes a camera for collecting image data, a displacement sensor for collecting displacement data, and a force sensor for collecting mechanical data, and integrates to form a preliminary measurement data set of the slump and spread of the concrete.

[0108] Specifically, the image data set of the concrete sample is obtained by the camera, and the image data set includes a plurality of image frames at different angles:

[0109] I = {I1, I2, …, In} n}

[0110] Where I i represents the i-th image frame, and n is the total number of frames.

[0111] The displacement data set of the concrete sample at different time points is measured by the displacement sensor:

[0112] D = {d1, d2, …, d m}

[0113] wherein d j represents the displacement value at the jth time point, and m is the total number of measurement time points;

[0114] The mechanical data set of the concrete sample under the loading condition is obtained by the force sensor:

[0115] F = {f1, f2, …, f k}

[0116] wherein f p represents the force value at the pth measurement point, and k is the total number of measurement points;

[0117] The image data set I, the displacement data set D and the mechanical data set F are integrated to form a preliminary measurement data set:

[0118] M = {I, D, F}.

[0119] Improvements are made on the feature layer, the loss layer and the weight update layer for the slump / extension multi-modal data:

[0120] On the feature layer, the image tensor I is encoded by sparse SIFT-SURF combined with gray gradient to obtain contour-texture, the displacement sequence D is constructed by time domain difference to obtain settlement velocity, and the mechanical sequence F is extracted by short-time Fourier transform to obtain cohesion-friction spectrum vector. The three features are aligned in the shared embedding space by mutual information maximization criterion.

[0121] On the loss layer, a coupling term is proposed and constructed, which directly corresponds to the physical quantity of the scene, to ensure that the distance metric learned by the model is simultaneously subject to the image contour, displacement curve and stress-rheological double constraints; the weight update layer adopts cross-validation directional search instead of fixed coefficients or simple voting, so that only single path data cannot meet the above joint optimization relationship and cannot achieve the same accuracy.

[0122] This improvement of the embodiment closely follows the specific technical problem of "partial marking + dynamic positioning" at the construction site: in real pouring, the sample state will decay over time, and non-uniform collapse may occur due to high temperature bleeding or aggregate accumulation. Simply relying on images may miss internal settlement, and relying solely on displacement cannot capture the radial expansion boundary, and missing mechanical information makes it more difficult to distinguish the difference in mixing degree. The improved model takes the coupling error of I, D and F as the optimization target, drives the robot to correct the positioning in real time and gives a consistent solution of height-diameter-force, thereby converting the abstract problem of incomplete data into the operable process of automatic resampling + modal compensation, and improving the consistency and repeatability of slump and spread on site.

[0123] Further based on the data usage problem, the three types of data are indispensable in the model training and inference process in this step: if only I is retained, the loss function degenerates into two-dimensional contour matching, and the force-displacement component no longer provides gradient information, and the gradient-step update rule is invalid; if only I and D are used without F, the coupling term disappears, and the extension rheological constraint is weakened, and the prediction will systematically underestimate under large aggregate ratio; even if other external data is introduced, it cannot participate in the mutual information alignment, nor can it enter the joint feature space. Therefore, the data of I, D and F are associated.

[0124] S4: After preprocessing the initial measurement data set, input it into the improved weakly supervised learning model for training, and output the prediction results of the concrete slump and spread.

[0125] Preferably, step S4 specifically includes the following steps:

[0126] S41: input the preprocessed measurement data set into the semi-supervised learning model integrated model based on the Scikit-learn framework for model training and verification, calculate the prediction accuracy of the model, and the Scikit-learn framework integrates semi-supervised k-nearest neighbor model and semi-supervised support vector machine model;

[0127] S42: use the semi-supervised k-nearest neighbor model and semi-supervised support vector machine model in the Scikit-learn framework to preliminarily train the measurement data set, respectively, to obtain their respective prediction results y kNN and y SVM ;

[0128] S43: integrate the output results of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model by voting to generate the final prediction result y final :

[0129] y final = α·y kNN + β·y SVM ; ​

[0130] wherein a and β are weight parameters of the corresponding model, and a + β = 1;

[0131] S44: In the training process of the integrated model, the weight parameters a and β of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model are dynamically adjusted through cross-validation, so that the integrated model obtains the best prediction performance under the condition of partial labeled data;

[0132] S45: Finally output the prediction results of the integrated model on the concrete slump and spread, and classify and predict the unlabeled data to generate the final evaluation value of the concrete slump and spread corresponding to the data.

[0133] Preferably, a unified metadata structure (X, y, δ) is established for each sample at the data access layer, wherein X represents a three-modal feature tensor, y is a target vector, and δ is a label state indicator. When the sample is measured by the artificial reference method at the same time, the slump h ref and the spread d ref , and the test error meets the specified ±2mm interval, write y = (h ref , d ref ) and δ = 1, if the reference value is missing, out of tolerance or still in the verification queue, y is assigned to 0, and δ is assigned to 0, realizing the physical meaning of labeled and unlabeled separation.

[0134] According to δ, the and are automatically constructed The former participates in the supervised part loss of semi-supervised kNN / SVM The latter enters the unsupervised part loss through consistency regularization and pseudo-label strategy

[0135] Specifically, S41 includes the following steps:

[0136] S411: Divide the preprocessed measurement data set into a training data set M train and a test data set M test , and the segmentation ratio is determined according to the following formula:

[0137]

[0138] wherein r is the proportion of the training set, λ1 is a proportion control parameter, σ M is the standard deviation of the measurement data set, is the reference standard deviation;

[0139] S412: Initialize the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework, and set the model parameters respectively. For the semi-supervised k-nearest neighbor model, the number of neighbors k is selected according to the following optimization:

[0140]

[0141] Where, η kNN To balance the parameters, Let i be the feature vector of the i-th training sample. For the corresponding label, |||| represents the Euclidean distance, and l(·,·) represents the label loss function;

[0142] For a semi-supervised support vector machine model, the kernel function parameter γ is determined through the following adaptive optimization:

[0143]

[0144] Among them, w SVM For the weight vector of the support vector machine, For the kernel function mapping of the input samples, μ SVM and ν SVM For regularization parameters, The output of the model using the kernel function parameter γ;

[0145] S413: Transfer the training dataset M train The inputs are fed into a semi-supervised k-nearest neighbor model and a semi-supervised support vector machine model, respectively, for training. The trained semi-supervised k-nearest neighbor model and semi-supervised support vector machine model are obtained. The training process is optimized by minimizing the following comprehensive loss function:

[0146]

[0147] Where, λ kNN , λ SVM λ3 are the weight parameters of the loss function. Let be the loss functions corresponding to the k-nearest neighbor model and the support vector machine model, respectively;

[0148] S414: Using the test dataset M test The trained semi-supervised k-nearest neighbor model and semi-supervised support vector machine model were validated, and the prediction accuracy of the models was calculated:

[0149]

[0150] in, As an indicator function, when predicting labels With real labels ω is 1 if they match, otherwise it is 0. kNN and ω SVM These are the weight coefficients for the k-nearest neighbor model and the support vector machine model, respectively.

[0151] Optionally, the loss function of the k-NN model and the SVM model are designed in combination with the prediction accuracy of the slump and the spread of concrete, the feature distance between samples, and the time series error.

[0152] The loss function of the k-NN model is as follows:

[0153]

[0154] wherein, represents the predicted slump or spread of concrete by the semi-supervised k-NN model, is the actual measured value, λ d1 is the weight coefficient for controlling the Euclidean distance between samples, λ d1 is the weight coefficient for controlling the time series error, is the feature vector of the i-th sample, and are the predicted and actual measured time, respectively;

[0155] The loss function of the SVM model is as follows:

[0156]

[0157] wherein, is the actual slump or spread label of concrete, is the predicted value generated by the SVM model, w SVM is the weight vector of the SVM, λ d2 is the regularization parameter for controlling the complexity of the model, λ t2 is the weight coefficient for controlling the time error, is the measured time predicted by the model, and b is the bias term.

[0158] Preferably, S42 comprises the following steps:

[0159] S421: input the feature data x i and the partially labeled target data to the semi-supervised k-NN model in the Scikit-learn framework, and preliminarily train the k-NN model according to the loss function of the k-NN model to obtain the prediction result y kNN of the k-NN model:

[0160]

[0161] wherein, w j is the weight of the j-th neighbor sample, y j is the predicted label of the sample, λ1 is the regularization parameter for adjusting the weight, x i and x jThese are the feature vectors of the sample to be tested and the neighboring samples, respectively.

[0162] S422: Transfer feature data from measurement dataset M and partially labeled target data The input is fed into a semi-supervised support vector machine (SVM) model within the Scikit-learn framework, and initial training is performed based on the SVM model's loss function to obtain the SVM model's prediction result y. SVM :

[0163]

[0164] Among them, w SVM For the weight vector of the support vector machine, λ is the kernel function mapping for the input features, b is the bias term, and λ2 is the regularization parameter.

[0165] S5: Based on the prediction results of the improved weakly supervised learning model, the robot system calculates the initial measurement position, parameter configuration, and adjusted measurement position and parameter configuration for each concrete sample. This yields the concrete slump and spread under different measurement positions and parameters, forming a sample dataset. The sample dataset is then analyzed, and the optimal measurement position and parameter configuration are determined based on the variance of the sample dataset. The final values ​​of concrete slump and spread, as well as the qualification result, are output accordingly.

[0166] Specifically, S5 includes the following steps:

[0167] A1: Output result y based on the improved weakly supervised learning model final The robot system calculates the initial measurement position P0(x0,y0,z0) of the concrete sample and its corresponding parameter configuration Θ0={θ1,θ2,…,θ n}, including the parameter configuration of the vision camera and laser used to acquire the position of the concrete sample, where P0 is the three-dimensional coordinate of the measurement position and Θ0 is the parameter set;

[0168] A2: The robot system automatically adjusts itself based on the initial measurement position P0 and parameter configuration Θ0, and calculates the adjusted measurement position P. adj (x adj ,y adj ,z adj ):

[0169] P adj =P0 + ΔP where

[0170] Where ΔP is the adjustment vector for the measurement position, and α3 is the adjustment step size. Represents the loss function for measurement location P. The gradient;

[0171] A3: Perform multiple measurements, and recalculate the adjusted measurement position P after each measurement. adj and the corresponding parameter configuration Θ adj To obtain sample datasets of concrete slump and spread under different measurement locations and parameters:

[0172]

[0173] A4: For sample dataset M meas The analysis was performed based on the variance Var(M) of the sample dataset. meas Determine the optimal measurement location P. opt and parameter configuration Θ opt Output the final concrete slump and spread values ​​y opt And the judgment result R opt :

[0174]

[0175] Where β3 is the weighting coefficient, y min and y max This is the preset range for determining whether a product is acceptable.

[0176] S6: Transmit the qualified values ​​determined in step S5 to the existing design-connected construction management system for visualization, and provide real-time feedback based on the measurement results to guide construction operations or adjust the concrete mix proportions.

[0177] The robot measurement system generally comprises a multi-section arm structure such as a mechanical large arm, a mechanical small arm, etc. which can be flexibly moved and rotated in space, a data acquisition module, a communication module, a control module for controlling the movement and positioning of the mechanical arm, a positioning module, etc., and the data acquisition module comprises a sensor module. In the Scikit-learn framework, the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model are integrated for the selected image data, displacement data and mechanical data of the concrete, thereby effectively solving the problem of insufficient processing capacity for incomplete labeled data in the prior art. By introducing the semi-supervised learning algorithm, the model can be trained under the condition of only having partial labeled data, thereby reducing the dependence on completely labeled data and improving the generalization ability of the model. Through the gradient optimization algorithm for the measurement position, the position and parameters are dynamically adjusted after each measurement, more accurate measurement data are obtained, the system determines the final optimal measurement position and parameter configuration according to the variance of the measurement results, and thus the most reliable slump and spread degree values are output, the automation degree of the measurement process is improved, the possibility of manual intervention is reduced, and the high precision and high repeatability of the measurement results are ensured.

[0178] Implementation case:

[0179] In order to verify the effectiveness of the present application, the present application is first applied to the QYTJ-2 project of Qin-I Expressway, the experimental time is June 15, 2025, the construction party needs to carry out concrete pouring work of bridge pier columns, the construction site is located in Gongyi, Henan, the construction environment is complex and changeable in high temperature and rainy summer, in order to ensure the engineering quality, the construction party decides to monitor the slump and spread degree of the concrete in real time during the concrete pouring process, and the monitoring data is used as a key basis for quality control.

[0180] On the morning of June 15, 2024, the construction party began the concrete pouring operation of the day, in order to verify the actual effect of the method of the present application, the construction party selected three key periods in the pouring process to monitor the slump and spread degree of the concrete, and the robot system was deployed at different positions in the construction site to automatically perform data acquisition and measurement tasks.

[0181] Firstly, the robot system measured the slump and spread degree of the first batch of concrete in the period from 9:00 to 10:30 am. The robot collected various measurement data including concrete surface images, displacement data during the slump process and mechanical data through the configured sensor module. These data were transmitted in real time to the semi-supervised learning model in the Scikit-learn framework through the data acquisition module.

[0182] In the next operation, the robot system initiates the training of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model according to the preliminary data obtained. In this process, the robot system utilizes the labeled data set and combines part of the unlabeled data set to generate the first batch of concrete slump and spread prediction results through the dynamic optimization mechanism of the model. According to the prediction results, the system automatically adjusts the measurement position and parameters, further optimizing the measurement accuracy.

[0183] The concrete sample is extracted by an integrated sampling unit connected to the end of the mechanical arm, which includes a telescopic sampling pipe, a screw propelling pump, a quick locking valve, and a discharging valve. The sampling pipe is inserted into the 150mm deep of the mixer truck hopper under the guidance of visual-laser double positioning, and the fluid concrete is extracted at a uniform speed of 150r / min and delivered to the conical mold placed on the detection platform. The mold volume is designed according to the standard slump test cone mold volume V c ≈5.5L, with a conical mold upper diameter of 100mm, a lower diameter of 200mm, and a height of 300mm. The upper and lower openings of the mold are equipped with linked switch valves. After completing the concrete feeding, the robot moves the mold at a uniform speed to the center position of the detection platform measurement reference, the lower valve of the mold is opened, and the robot lifts the mold upward linearly within 3-7 seconds, and the concrete in the mold naturally collapses.

[0184] The mechanical data is obtained by a six-axis force-torque sensor installed below the measurement table. The sensor has a range of ±5kN and a resolution of 0.5N. The static calibration uses traceable weights to F z The direction is loaded in ten grades, and the linear fitting slope error is controlled within 0.18%. The dynamic characteristics are measured by step excitation, and the first-order natural frequency is 320Hz, which is more than three times the main frequency of the slump process, and can be considered as a quasi-static process. The displacement sensor is a combination of laser displacement meter and visual structured light. The laser channel is used to monitor the center point settlement δ h (t), and the structured light is used to reconstruct the surface grid and calculate the radial diffusion diameter d(t). The two sets of sensors are synchronized under the same PTP clock, with a time drift of <1ms, ensuring the consistency of the time stamp of the force-displacement-image three-modal data, providing an accurate reference for subsequent weakly supervised feature fusion.

[0185] The adaptive update of the robot measurement pose is realized by minimizing the prediction residual loss , where p is the three-dimensional pose, θ is the visual camera and laser parameter. The post-measurement residual e i is approximated by finite difference , and the step size λ=3mm is updated θ is iterated between exposure, focal length, and structured light projection angle using the dual variable method, with the constraint condition that the picture saturation is <5% and the feature corner contrast is >0.4. If |y i+1 -yi | / y i <0.5% is determined to be converged and the final slump h is output * and the spread d * .

[0186] To verify the effectiveness of the method of the application, the construction party compared the method of the application with the traditional manual measurement method of the field laboratory. The following Table 1 is the specific measurement data and results:

[0187] Table 1 Comparison of the method of the application and the traditional manual measurement method

[0188]

[0189] Through comparison, it can be seen that the benchmark values of the slump and the spread measured manually are 225 mm and 390 mm respectively (note: usually, the data processing of the slump and the spread measured by the traditional manual method on the construction site usually adopts the "rounding method with 5 as the base", that is, the value is usually close to the nearest multiple of 5, so as to facilitate rapid statistics, and the data processing is relatively arbitrary, and the measurement accuracy is not high). The measurement value of the algorithm of the application remains at a low level in multiple tests, the error of the slump is between 0.68% and 1.67%, and the error of the spread is between 0.99% and 1.84%, the error is small and very stable, which shows that the method of the application can still maintain high accuracy when dealing with complex construction environment.

[0190] The algorithm measurement result of the method of the application shows very stable error control, especially in multiple measurements, the error range is less than 2%, compared with the traditional manual test measurement on the site, which needs to be completed by 2 professional laboratory personnel, not only the efficiency is low, but also it is easily affected by the experience of the operator, the construction environment and other factors, and the error can be larger.

[0191] In multiple measurements, the algorithm measurement result of the application shows high consistency, the standard deviation of the slump in multiple measurements is less than 1 mm, and the standard deviation of the spread is also controlled within 1.5 mm, such high consistency effectively avoids data fluctuation caused by environment or human factors, and improves the reliability of the measurement result.

[0192] The traditional manual measurement needs more than 20 minutes each time (including the time of connecting the material to the discharge port of the mixing truck, feeding, and manual test measurement), and the automatic measurement system based on the application can complete one complete slump and spread measurement in about 10 minutes, saving at least half of the time, improving the work efficiency of the construction site, especially in large-scale construction projects, the automatic measurement can greatly improve the work progress.

[0193] Through the above comparison, the robot concrete slump and spread measurement method based on improved weak supervision learning has superior performance in measurement accuracy and efficiency. The system can effectively reduce human error in the measurement process, and still achieve high-precision prediction results through a semi-supervised learning model under incomplete labeled data. The application is suitable for various complex construction environments, especially in large-scale engineering projects, and can significantly improve the quality control level and construction efficiency of concrete measurement.

[0194] Although embodiments of the application have been disclosed in connection with the above specification and drawings, it should be understood that they are not limited to the particular form it is disclosed but includes all modifications and equivalents of the concepts described. It is intended to cover by the claims any and all modifications and uses of the application within the true spirit and scope of the application.

Claims

1. A method for measuring slump and spread of concrete by a robot based on weakly supervised learning, characterized in that, Comprising the following steps: S1: setting and initializing the robot system, configuring the connection sensor module for acquiring image data, displacement data, and mechanical data of the concrete to be measured; S2: obtaining the concrete sample to be measured, placing the concrete sample to be measured at a designated position for measurement by the robot system, and automatically adjusting the concrete sample to a preset reference position with a deviation value not exceeding the set deviation value from the reference position; S3: starting the sensor module to collect data and integrating to form a preliminary measurement data set of the concrete slump and spread; S4: after preprocessing the preliminary measurement data set, inputting it into the improved weakly supervised learning model for training, and outputting the prediction results of the concrete slump and spread; S5: based on the prediction results of the improved weakly supervised learning model, calculating the initial measurement position, parameter configuration, and adjusted measurement position, parameter configuration of the concrete sample each time through the robot system, obtaining the concrete slump and spread under different measurement positions and parameters, forming a sample data set, analyzing the sample data set, determining the optimal measurement position and parameter configuration according to the variance of the sample data set, and corresponding outputting the final numerical value of the concrete slump and spread and the determination result of whether it is qualified; S6: visualizing and displaying the numerical value determined to be qualified in step S5.

2. The robotic concrete slump and spread measurement method of claim 1, wherein, The sensor module includes a vision camera for collecting image data, a displacement sensor for collecting displacement data, and a force sensor for collecting mechanical data, wherein the vision camera is arranged towards the concrete sample, the force sensor is installed below the measurement table, and the displacement sensor is a combination of a laser displacement meter and a vision structured light, the laser channel is used to monitor the settlement of the center point, and the structured light is used to reconstruct the surface grid and calculate the radial diffusion diameter.

3. The robotic concrete slump and spread measurement method of claim 2, wherein, Step S2 specifically comprises the following steps: S21: After obtaining the concrete sample to be measured, the robot system automatically grabs the concrete sample through the mechanical arm, and moves the concrete sample to the designated position for measurement at a constant speed v c The designated position and the deviation value ∈ p are pre-set according to the measurement task, align the center of the concrete sample with the reference point of the measuring device; S22: measure the actual position coordinates P(x, y, z) of the concrete sample using the displacement sensor s (x s ,y s ,z s ) of the concrete sample and the preset reference position coordinates P r (x r ,y r ,z r ) Wherein, ΔP represents the deviation value of the actual position of the concrete sample from the reference position; S23: When ΔP > ε p , the position of the concrete sample is fine-tuned by the robotic system until ΔP ≤ ε p ; S24: after fine tuning, fixing the position of the concrete sample by the robot system.

4. The robotic concrete slump and spread measurement method of claim 2, wherein, In the step S3: Obtain the image data set of the concrete sample through the camera, which includes multiple image frames at different angles: I = {I1, I2,..., In} is a set of n images. n} where I i denotes the i-th image frame, and n is the total number of frames. Measure the displacement data set of the concrete sample at different time points using the displacement sensor: D = {d1, d2,..., d m} wherein d j denotes the displacement value at the jth time point, and m is the total number of measurement time points; Obtain the mechanical data set of the concrete sample under the loading condition through the force sensor: F = {f1, f2,..., f k} wherein f p represents the force value of the pth measurement point, and k is the total number of measurement points; Integrate the image data set I, displacement data set D, and mechanical data set F to form a preliminary measurement data set: M={I,D,F}.

5. The robotic concrete slump and spread measurement method of claim 1, wherein, The step S4 specifically comprises the following steps: S41: input the preprocessed measurement data set into the semi-supervised learning model integration model based on the Scikit-learn framework for model training and verification, calculate the prediction accuracy of the model, and integrate the semi-supervised k-nearest neighbor model and semi-supervised support vector machine model in the Scikit-learn framework; S42: Use the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework to preliminarily train the measurement data set respectively to obtain respective prediction results y kNN and y SVM ; S43: Integrate the output results of the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model by voting to generate the final prediction result y final : y final = a · y kNN + b · y SVM ; Wherein, α and β are the weight parameters of the corresponding model, and α+β=1; S44: during the training process of the integrated model, dynamically adjust the weight parameters α and β of the semi-supervised k-nearest neighbor model and semi-supervised support vector machine model through cross-validation, so that the integrated model obtains the best prediction performance under the condition of partial labeled data; S45: Finally output the prediction results of the integrated model on the concrete slump and spread, and classify and predict the unlabeled data to generate the final evaluation value of the corresponding data of the concrete slump and spread.

6. The robotic concrete slump and spread measurement method of claim 5, wherein, S41 includes the following steps: S411: Split the pre-processed measurement dataset into a training dataset M train and a test dataset M test The split ratio is determined according to the following: wherein r is the proportion of the training set, λ1is a proportion control parameter, σ M is the standard deviation of the measurement data set, is the reference standard deviation; S412: Initialize the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model in the Scikit-learn framework, and set the model parameters respectively. For the semi-supervised k-nearest neighbor model, the selection of the number of neighbors k is based on the following optimization: where η kNN is a trade-off parameter, is the feature vector of the i-th training sample, is the corresponding label, and ||·|| denotes the Euclidean distance, l(·, ·) denotes the label loss function. For the semi-supervised support vector machine model, the selection of the kernel function parameter γ is determined by the following adaptive optimization: where w SVM is a weight vector of the support vector machine, is a kernel function mapping of the input sample, μ SVM and v SVM are regularization parameters, is a model output using a kernel function parameter γ; S413: input the training data set M train into the semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model, respectively, to perform model training, to obtain the trained semi-supervised k-nearest neighbor model and the semi-supervised support vector machine model, and the training process is optimized by minimizing the following comprehensive loss function: wherein, λ kNN , λ SVM , λ3is a weight parameter of the loss function, are loss functions corresponding to the k-nearest neighbor model and the support vector machine model, respectively; S414: Using the test dataset M test The trained semi-supervised k-nearest neighbor model and semi-supervised support vector machine model are validated, and the prediction accuracy of the model is calculated: where, is an indicator function that is 1 when the predicted label is consistent with the true label and 0 otherwise, and ω kNN and ω SVM are the weight coefficients for the k- nearest neighbor model and the support vector machine model, respectively.

7. The robotic concrete slump and spread measurement method of claim 6, wherein, The loss functions of the k-nearest neighbor model and the support vector machine model are designed in combination with the prediction accuracy of the concrete slump and spread, the feature distance between samples, and the time series error; The loss function of the k-nearest neighbor model is as follows: wherein, represents the concrete slump or spread predicted by the semi-supervised k-nearest neighbor model, is the actual measured value, λ d1 is the weight coefficient to control the Euclidean distance between samples, λ d1 is the weight coefficient to control the time series error, is the feature vector of the i-th sample, and are the model predicted and real measured time, respectively; The loss function of the support vector machine model is as follows: wherein, is the true concrete slump or spread label, is the predicted value generated using the support vector machine model, w SVM is the weight vector of the support vector machine, λ d2 is the regularization parameter that controls the complexity of the model, λ t2 is the weight coefficient that controls the time error, is the measured time, b is the bias term.

8. The robotic concrete slump and spread measurement method of claim 6, wherein, S42 includes the following steps: S421: Feature data x in the measurement dataset is inputted into the semi-supervised k- nearest neighbor model in the Scikit-learn framework, and preliminary training is performed according to a k-nearest neighbor model loss function to obtain a prediction result y of the k-nearest neighbor model i and partially labeled target data into the semi-supervised k-nearest neighbor model in the Scikit-learn framework, and preliminary training is performed according to a k-nearest neighbor model loss function to obtain a prediction result y of the k-nearest neighbor model kNN : where w j is the weight of the jth neighbor sample, y j is the predicted label of the sample, λ1is the regularization parameter of the adjustment weight, x i and x j are the feature vectors of the sample to be tested and the neighbor sample, respectively; S422: inputting the feature data in the measurement data set M and the target data in the partially labeled target data set T into a semi-supervised support vector machine model in a Scikit-learn framework, and performing preliminary training according to a support vector machine model loss function to obtain a prediction result y of the support vector machine model and the partially labeled target data input into the semi-supervised support vector machine model in the Scikit-learn framework, and performing preliminary training according to a support vector machine model loss function to obtain a prediction result y of the support vector machine model SVM : where w SVM is the weight vector of the support vector machine, is the kernel function mapping of the input features, b is the bias term, and λ2is the regularization parameter.

9. The robotic concrete slump and spread measurement method of claim 6, wherein, When labeling data, a unified metadata structure (X, y, δ) is established for each sample at the data access layer, where X represents the three-modal feature tensor, y is the target vector, and δ is the labeling status indicator. When the slump h of the sample is simultaneously measured using the manual reference method... ref With expansion degree d ref And when the verification error meets the specified ±2mm range, write y=(h ref ,d ref And set δ to 1. If the benchmark value is missing, out of tolerance, or still in the verification queue, then assign y to... Setting δ to 0 achieves physical separation between labeled and unlabeled flows. auto-constructing per delta and the former participates in the supervised part loss of semi-supervised kNN / SVM the latter enters the unsupervised part loss with consistency regularization and pseudo-label strategy 10. The robotic concrete slump and spread measurement method of claim 1, wherein, S5 specifically includes the following steps: A1: output results y based on an improved weakly supervised learning model final , calculate the initial measurement position P0(x0, y0, z0) of the concrete sample and its corresponding parameter configuration Θ0={θ1, θ2, …, θ n} by the robot system, including the parameter configuration of the vision camera and the laser for obtaining the position of the concrete sample, wherein P0 is the three-dimensional coordinates of the measurement position, and Θ0 is the parameter set; A2: The robot system automatically adjusts according to the initial measurement position P0 and the parameter configuration Θ0, and obtains the adjusted measurement position P by calculation adj (x adj ,y adj ,z adj ) P adj = P0+ ΔP where wherein ΔP is an adjustment vector of the measurement position, and a3 is an adjustment step size, denotes the gradient of the loss function with respect to the measurement position P; A3: performing multiple measurements, after each measurement, recalculating the adjusted measurement position P adj and the corresponding parameter configuration Θ adj obtaining a sample dataset of concrete slump and spread at different measurement positions and parameters: A4: For the sample data set M meas analysis, according to the variance Var(M meas ), determine the optimal measurement position P opt and parameter configuration Θ opt , output the final concrete slump and spread of the numerical value y opt and the determination result R opt : wherein β3 is a weight coefficient, y min and y max is a preset qualification determination range.

Images