Concrete mixing monitoring and fiber dispersibility evaluation method based on data fusion

Abstract

The invention discloses a concrete mixing monitoring and fiber dispersibility evaluation method based on data fusion. According to the method, the scene adaptability and the evaluation accuracy of the whole monitoring method are remarkably improved by adaptively adjusting the contribution weight of the video dynamic flow value and the power fluctuation variance in the whole stirring period. According to the algorithm, feature weights are dynamically distributed innovatively according to the stirring process, a higher weight is given to a dynamic flow value reflecting macroscopic fluidity in the material mixing stage at the initial stirring stage, and the flow features of preliminary fusion of fibers and a matrix are mainly captured; and entering a stable period along with stirring, automatically transitioning to evaluation logic dominated by power fluctuation variance, and accurately identifying the microscopic stable state after the fibers are uniformly dispersed. The dynamic adjustment mechanism effectively solves the problem of evaluation deviation of traditional fixed weight fusion in different stirring stages, and further enhances the evaluation robustness under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of steel fiber mixing monitoring technology, specifically a method for concrete mixing monitoring and fiber dispersion evaluation based on data fusion. Background Technology

[0002] Ultra-high performance concrete (UHPC) is widely used due to its superior mechanical properties and durability. The incorporation of steel fibers can significantly improve the toughness and tensile strength of UHPC. However, the uniformity of steel fiber dispersion in the matrix is ​​a key factor affecting the final performance of UHPC. Fiber clumping or uneven distribution can create weak areas, leading to a sharp decline in performance.

[0003] Currently, the evaluation of fiber dispersibility relies heavily on ex-post, destructive testing methods, such as: 1. Post-hardening sectional method: The hardened specimen is sectionalized, and the fiber dispersion coefficient (such as the number of fibers per unit area and fiber orientation) is calculated through image processing. This method has a lag effect and cannot provide real-time feedback and adjustment to the production process.

[0004] 2. Experience-based judgment: Relies on workers' experience in observing the state of the mixed material, which is highly subjective and lacks quantitative standards.

[0005] Therefore, existing technologies lack an effective means to monitor and evaluate the uniformity of steel fiber dispersion in real time, online, and non-destructively during the mixing process. Summary of the Invention

[0006] The purpose of this invention is to provide a method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion in order to solve the problems mentioned above.

[0007] The technical solution adopted in this invention is as follows: a method for monitoring concrete mixing and evaluating fiber dispersibility based on data fusion, comprising the following steps: S1: An industrial camera is fixed in the observation window or on top of the mixer to record video of material flow during the mixing process at a fixed frame rate; real-time power data of the motor is synchronously collected through a power sensor, with a sampling frequency ≥10Hz. The video data is preprocessed with anti-shake and noise reduction, and the power data is subjected to outlier removal and smoothing to ensure that the data quality meets the requirements of the subsequent flowability assessment in S2 and the stability analysis in S3.

[0008] S2: The preprocessed video stream from S1 is input into the trained flowability assessment model. This model is trained based on historical mixing video samples and corresponding slump flow value labels, and can output dynamic flow values ​​that reflect the macroscopic flow state of the material. The dynamic flow values ​​of the current 30-second video window are calculated in real time, and the results, together with the power fluctuation characteristics output from S3, will serve as the fusion basis for S4.

[0009] S3: Extract the time series of the power data collected in S1 during the later stable stage of stirring (20-30 seconds after feeding), and calculate the power fluctuation variance during this period. Experimental verification shows that the variance value is negatively correlated with the fiber dispersion uniformity. This characteristic value will be input into S4 along with the dynamic flow value from S2 for comprehensive evaluation.

[0010] S4: Based on the dynamic flow value output from S2 and the power fluctuation variance extracted from S3, a dynamic weighted fusion model is constructed. The weights of the two indicators are automatically adjusted according to the stirring process, and the fiber dispersion uniformity index is obtained through weighted calculation. A threshold is set by combining historical qualified sample data. When the index is lower than the threshold, an alarm is triggered to achieve real-time detection of aberrations in stirring quality.

[0011] In a preferred embodiment, in step S1, video data acquisition requires fixing an industrial camera with a resolution of no less than 1920×1080 at the observation window or top center of the mixer. The lens focal length is adjusted according to the diameter of the mixing drum to ensure complete coverage of the material flow area. The frame rate is set to 25 frames per second, and the recording period is from the start of feeding until 10 seconds after the end of mixing, focusing on capturing the dynamic process of material tumbling and flowing after feeding. Simultaneously, power data acquisition is initiated using a current-voltage combined power sensor with an accuracy class of 0.5. The motor power signal is acquired in real time through the inverter's analog output interface, and the sampling frequency must reach at least 10 Hz to ensure that power fluctuation details at the 0.1-second level during fiber dispersion can be captured.

[0012] In a preferred embodiment, in step S1, during the data preprocessing stage, the video data is first enhanced with grayscale-based contrast to eliminate brightness unevenness caused by flickering stirring lights. Then, Gaussian filtering is used to remove dust noise from the industrial environment. Finally, optical flow is used for adjacent frame registration to eliminate image shift caused by slight camera vibration. Power data preprocessing uses the 3σ criterion to remove peak outliers at the moment of motor startup, followed by smoothing using a moving average filter with a sliding window size of 50 sampling points to ensure that data fluctuation characteristics are not affected by high-frequency noise. The processed video and power data are stored as timestamp-aligned data streams, providing synchronous input for the flow assessment in S2 and the stability analysis in S3.

[0013] In a preferred embodiment, in step S2, the model training phase requires constructing a sample library containing more than 200 sets of different working conditions. Each set of samples covers UHPC stirring videos with fiber volume content of 1% to 3% and water-to-binder ratio of 0.2 to 0.3. Simultaneously, the 30-second flow value of the corresponding sample is measured using a collapse spread test as a label, with the flow value range controlled between 150 mm and 250 mm. A hybrid architecture of 3D convolutional neural network and long short-term memory network is adopted as the flow evaluation model. The 3DCNN layer extracts spatial flow features from the video clips, and the LSTM layer captures the flow trend changes in the time dimension. The model input is a continuous 150-frame (6-second) video clip, and the output is a dynamic flow value (DFV) normalized to 0 to 100. The mean squared error loss function is used during training, and the iteration is continued until the validation set error is less than 5 mm.

[0014] In a preferred embodiment, in step S2, during online application, the system captures the video stream (750 frames) up to 30 seconds prior to the current moment in real time, divides it into 5 consecutive segments using a 6-second sliding window, and inputs each segment into a pre-trained model to obtain 5 intermediate DFV values. The final real-time DFV is calculated by weighted averaging, with the weight allocation dynamically adjusted according to the mixing process. Within 10 seconds after material addition, the weight is biased towards later segments to reduce interference in the initial mixing stage. This DFV value is transmitted in real time to S4, where it, along with the PFV output from S3, participates in the calculation of the dispersion uniformity index.

[0015] In a preferred embodiment, in step S3, the power data extraction first requires determining the starting point of the stable phase using a stirring process identification algorithm. The stable phase begins when the absolute value of the power change rate is less than 0.5% for 5 consecutive seconds, typically corresponding to the 20-30 second interval after feeding. The extraction duration is fixed at 20 seconds to ensure data consistency. Feature extraction is then performed on the extracted power time series. First, the average power μ within that time period is calculated, and then the power is calculated using the formula PFV=Σ(pi-μ). 2 / (n-1) calculates the fluctuation variance, where pi is the power value of the i-th sampling point, n is the number of sampling points (20 seconds × 10 Hz = 200 points), and PFV is in watts squared, reflecting the degree of dispersion of power from the mean per unit time.

[0016] In a preferred embodiment, step S4 specifically includes the following process: S4.1 Data Spatiotemporal Alignment: The dynamic flow value (DFV) output from step S2 and the power fluctuation variance (PFV) extracted from step S3 in the same mixing batch are spatiotemporally matched to ensure that they correspond to the same mixing stage (e.g., the stabilization period data 20-30 seconds after the feeding is completed). Data alignment is achieved through timestamp synchronization to provide consistent input for subsequent fusion calculations. S4.2 Linear weighted fusion calculation: The fiber dispersion uniformity index DUI is constructed using a linear weighting method, as shown in the following formula: DUI = a·DFV−b·PFV+c; in: DFV is the dynamic flow value output in step S2, which reflects the macroscopic flowability of the material. The higher the value, the more uniform the flow state. PFV is the power fluctuation variance calculated in step S3, which reflects the stirring stability. The lower the value, the more uniform the fiber dispersion. a, b, and c are coefficients fitted by regression of experimental data. They were obtained by collecting DFV, PFV, and corresponding fiber dispersion coefficients of hardened sections from more than 100 qualified samples and solving them using the least squares method.

[0017] S4.3 Dynamic Threshold Determination and Feedback: A pass / fail threshold is set based on the statistical values ​​of DUI from historical qualified batches (e.g., the lower limit of the 95% confidence interval). After real-time calculation of the current batch's DUI, if it falls below the threshold, the system automatically triggers an alarm, prompting the operator to adjust mixing parameters (e.g., extending the mixing time or optimizing the feeding sequence). The DUI data for this batch is then fed back to steps S2 and S3 to update the sample library for subsequent model training.

[0018] In a preferred embodiment, in step S4.1, to eliminate dimensional differences and unify the data range, the dynamic flow value DFV output in S2 and the power fluctuation variance PFV extracted in S3 are standardized, wherein the DFV standardization formula is: ; The PFV standardization formula is: ; In the formula: DFV represents the dynamic flow value output by S2 (original range: 0~200mm, corresponding to UHPC slump flow value). DFV min / DFV max This represents the minimum / maximum DFV value in the historical samples (determined through statistical analysis of 100 sets of experiments, e.g., DFV). min =50mm, DFV max =180mm); PFV represents the power fluctuation variance calculated by S3 (original unit: W). 2 Range: 0.5~15 W 2 ); PFV min / PFV max This represents the minimum / maximum PFV value in historical samples (e.g., PFV). min =0.8 W 2PFV max =12W 2 ); DFV norm / PFV norm This represents the standardized feature value, with a uniform range of [0,1]. Higher values ​​indicate better liquidity and stronger dispersion stability, respectively.

[0019] In a preferred embodiment, in step S4.2, the DFV is adaptively adjusted based on the stirring process, with the time t after the feeding is completed as the variable. norm With PFV norm The weighting is as follows: In the initial stirring stage (t≤30s), material flowability dominates the dispersion process; in the later stage (t>30s), power stability better reflects the fiber dispersion state. The weighting function formula is as follows: ; ; In the formula: t represents the current mixing time (unit: s, starting from the moment the material is fed in). t0 represents the time of the weight inflection point (the experiment determined it to be 30s, meaning that after 30s, the stability weight exceeds the liquidity weight). k represents the weight adjustment rate coefficient (ranging from 0.1 to 0.3, with k=0.2 resulting in a smoother weight transition). w1(t) / w2(t) represents DFV norm / PFV norm The dynamic weights transition smoothly from 1 to 0 and from 0 to 1 as t progresses (w1=1, w2=0 when t=0; w1=0, w2=1 when t→∞).

[0020] In a preferred embodiment, in step S4.3, standardized features and dynamic weights are integrated to construct a comprehensive evaluation index, as shown in the following formula: In the formula: Δ(t) represents the stage compensation coefficient (which is linearly adjusted with t, Δ=0.1 when t=0 and Δ=0.05 when t=60s), used to offset the influence of data noise in the initial stage of stirring; DUI(t) represents the real-time dispersion uniformity index (range: 0~1.2), and the higher the value, the better the dispersion (the historical average DUI value for qualified samples is 0.85±0.1).

[0021] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, by adaptively adjusting the contribution weights of the video dynamic flow value and power fluctuation variance throughout the entire stirring cycle, the scene adaptability and evaluation accuracy of the entire monitoring method are significantly improved. This algorithm innovatively dynamically allocates feature weights according to the stirring process. In the initial material mixing stage, a higher weight is given to the dynamic flow value, which reflects macroscopic fluidity, focusing on capturing the flow characteristics of the initial fusion of fibers and matrix. As stirring enters the stable period, it automatically transitions to an evaluation logic dominated by power fluctuation variance, accurately identifying the microscopic stable state after fiber dispersion. This dynamic adjustment mechanism effectively solves the evaluation deviation problem of traditional fixed-weight fusion at different stages of stirring. Simultaneously, by standardizing features to unify dimensional differences, combined with the mixing stage compensation coefficient and dynamic threshold design, the evaluation robustness under complex working conditions is further enhanced. In practical applications, it improves the real-time calculation accuracy of the fiber dispersion uniformity index, enabling more sensitive identification of early fiber agglomeration risks and providing a reliable basis for timely adjustment of stirring parameters.

[0022] 2. In this invention, by collecting video streams and motor power data in real time during the mixing process, dynamic monitoring of fiber dispersion can be achieved simultaneously with material mixing. When the system detects abnormal fiber dispersion, it can immediately trigger an early warning, allowing operators to adjust mixing parameters on-site, such as extending the mixing time or optimizing the feeding sequence, thus avoiding the scrapping of the entire batch of materials and significantly improving the controllability and efficiency of the production process.

[0023] By fusing visual and electrical data, the macroscopic flow characteristics reflected by video data and the microscopic interaction characteristics reflected by power data complement each other. When the material flow is uniform and the power fluctuation is stable, good fiber dispersion can be cross-validated. If the flow is abnormal but the power is stable, or vice versa, the system can investigate the root cause of the problem through data correlation. This multi-dimensional verification mechanism greatly reduces the probability of misjudgment by a single indicator, making the evaluation results closer to the actual fiber dispersion state.

[0024] 3. In this invention, after training with a large number of labeled samples, the model can automatically extract flow features from the video, such as vortex morphology and material diffusion speed, and output quantitative flow indicators without human intervention. At the same time, the fluctuation analysis of power data is also automatically completed by the algorithm. The entire monitoring process, from data acquisition to result output, is fully automated, which avoids human error and reduces reliance on highly skilled workers, enabling the production line to maintain consistent monitoring accuracy.

[0025] 4. In this invention, non-contact monitoring is achieved through an industrial camera and a power sensor. The camera is installed outside the mixing tank, and power data is collected indirectly through the sensor. The entire process does not involve physical contact with the material, does not change the original state of the material, and does not affect the normal operation of the mixing equipment. This non-invasive design allows monitoring to be carried out synchronously with the production process without interrupting production or sacrificing test specimens, significantly improving the feasibility of industrial applications.

[0026] 5. In this invention, when applied to glass fiber reinforced cement, only the flowability characteristic parameters in the video model need to be adjusted to adapt to the flow characteristics of the cement, or the power fluctuation threshold needs to be modified to match the interaction strength between the glass fiber and the matrix. For carbon fiber reinforced plastics, the camera's spectral range can be optimized to meet the imaging requirements of dark materials. This flexibility allows the method to be reused across material types, providing a universal solution for the mixing quality control of different fiber-reinforced composite materials, expanding the application scenarios and industrial value of the technology. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Example: Reference Figure 1 A data fusion-based method for monitoring concrete mixing and evaluating fiber dispersion includes the following steps: S1: An industrial camera is fixed in the observation window or on top of the mixer to record video of material flow during the mixing process at a fixed frame rate; real-time power data of the motor is synchronously collected through a power sensor, with a sampling frequency ≥10Hz. The video data is preprocessed with anti-shake and noise reduction, and the power data is subjected to outlier removal and smoothing to ensure that the data quality meets the requirements of the subsequent flowability assessment in S2 and the stability analysis in S3.

[0030] S2: The preprocessed video stream from S1 is input into the trained flowability assessment model. This model is trained based on historical mixing video samples and corresponding slump flow value labels, and can output dynamic flow values ​​that reflect the macroscopic flow state of the material. The dynamic flow values ​​of the current 30-second video window are calculated in real time, and the results, together with the power fluctuation characteristics output from S3, will serve as the fusion basis for S4.

[0031] S3: Extract the time series of the power data collected in S1 during the later stable stage of stirring (20-30 seconds after feeding), and calculate the power fluctuation variance during this period. Experimental verification shows that the variance value is negatively correlated with the fiber dispersion uniformity. This characteristic value will be input into S4 along with the dynamic flow value from S2 for comprehensive evaluation.

[0032] S4: Based on the dynamic flow value output from S2 and the power fluctuation variance extracted from S3, a dynamic weighted fusion model is constructed. The weights of the two indicators are automatically adjusted according to the stirring process, and the fiber dispersion uniformity index is obtained through weighted calculation. A threshold is set by combining historical qualified sample data. When the index is lower than the threshold, an alarm is triggered to achieve real-time detection of aberrations in stirring quality.

[0033] In step S1, video data acquisition requires fixing an industrial camera with a resolution of no less than 1920×1080 at the observation window or top center of the mixer. The lens focal length should be adjusted according to the diameter of the mixing drum to ensure complete coverage of the material flow area. The frame rate should be set to 25 frames per second, and the recording period should be from the start of feeding until 10 seconds after the mixing ends, focusing on capturing the dynamic process of material tumbling and flowing after feeding. Simultaneously, power data acquisition should be started using a current-voltage combined power sensor with an accuracy class of 0.5. The motor power signal should be acquired in real time through the inverter's analog output interface, and the sampling frequency should reach at least 10 Hz to ensure that power fluctuation details at the 0.1-second level during fiber dispersion can be captured.

[0034] In step S1, during the data preprocessing stage, the video data is first enhanced with grayscale-based contrast to eliminate brightness unevenness caused by flickering stirring lights. Then, Gaussian filtering is used to remove dust noise from the industrial environment. Finally, optical flow is used for adjacent frame registration to eliminate image shift caused by slight camera vibration. Power data preprocessing uses the 3σ criterion to remove peak outliers at the moment of motor startup, followed by smoothing using a moving average filter with a sliding window size of 50 sampling points to ensure that data fluctuation characteristics are not affected by high-frequency noise. The processed video and power data are stored as time-stamp-aligned data streams, providing synchronous input for the flow assessment in S2 and the stability analysis in S3.

[0035] In step S2, during the model training phase, a sample library containing more than 200 different working conditions needs to be constructed. Each sample includes UHPC stirring videos with fiber volume content of 1% to 3% and water-cement ratio of 0.2 to 0.3. Simultaneously, the 30-second flow value of the corresponding sample is measured as a label through a collapse spread test.

[0036] The acquisition of the dataset specifically includes: recording the entire mixing process and simultaneously monitoring power data; measuring flowability through slump tests; and obtaining corresponding fiber dispersion data by cutting the hardened specimens and performing image processing. This process yields the optimal flowability for fiber dispersion. Then, image processing combined with power analysis is used for learning. The image identifies the slurry surface to ensure good fiber dispersion flowability, while the power analysis reflects the internal uniformity during mixing.

[0037] The flow value range is controlled between 150 mm and 250 mm. A hybrid architecture of 3D convolutional neural network and long short-term memory network is used as the flow evaluation model. The 3D CNN layer extracts spatial flow features from video clips, and the LSTM layer captures the flow trend changes in the temporal dimension. The model input is a continuous 150-frame (6-second) video clip, and the output is a dynamic flow value (DFV) normalized to 0 to 100. The mean squared error loss function is used during training, and the iteration is until the error on the validation set is less than 5 mm.

[0038] In step S2, during online application, the system captures the video stream of the previous 30 seconds (750 frames) in real time, divides it into 5 consecutive segments using a 6-second sliding window, and inputs each segment into a pre-trained model to obtain 5 intermediate DFV values. The final real-time DFV is calculated by weighted averaging, with the weights dynamically adjusted according to the mixing process. Within 10 seconds of material addition, the weights are biased towards later segments to reduce interference in the initial mixing stage. This DFV value is transmitted in real time to S4, where it, along with the PFV output from S3, participates in the calculation of the dispersion uniformity index.

[0039] In step S3, power data extraction first requires determining the starting point of the stable phase using a stirring process identification algorithm. The stable phase begins when the absolute value of the power change rate is less than 0.5% for 5 consecutive seconds, typically corresponding to 20 to 30 seconds after the material feeding is completed. The extraction duration is fixed at 20 seconds to ensure data consistency. Feature extraction is then performed on the extracted power time series. First, the average power μ within that time period is calculated, and then the formula PFV=Σ(pi-μ) is used. 2 / (n-1) calculates the fluctuation variance, where pi is the power value of the i-th sampling point, n is the number of sampling points (20 seconds × 10 Hz = 200 points), and PFV is in watts squared, reflecting the degree of dispersion of power from the mean per unit time.

[0040] Step S4 includes the following specific procedures: S4.1 Data Spatiotemporal Alignment: The dynamic flow value (DFV) output from step S2 and the power fluctuation variance (PFV) extracted from step S3 in the same mixing batch are spatiotemporally matched to ensure that they correspond to the same mixing stage (e.g., the stabilization period data 20-30 seconds after the feeding is completed). Data alignment is achieved through timestamp synchronization to provide consistent input for subsequent fusion calculations. S4.2 Linear weighted fusion calculation: The fiber dispersion uniformity index DUI is constructed using a linear weighting method, as shown in the following formula: DUI = a·DFV−b·PFV+c; in: DFV is the dynamic flow value output in step S2, which reflects the macroscopic flowability of the material. The higher the value, the more uniform the flow state. PFV is the power fluctuation variance calculated in step S3, which reflects the stirring stability. The lower the value, the more uniform the fiber dispersion. a, b, and c are coefficients fitted by regression of experimental data. They were obtained by collecting DFV, PFV, and corresponding fiber dispersion coefficients of hardened sections from more than 100 qualified samples and solving them using the least squares method.

[0041] S4.3 Dynamic Threshold Determination and Feedback: A pass / fail threshold is set based on the statistical values ​​of DUI from historical qualified batches (e.g., the lower limit of the 95% confidence interval). After real-time calculation of the current batch's DUI, if it falls below the threshold, the system automatically triggers an alarm, prompting the operator to adjust mixing parameters (e.g., extending the mixing time or optimizing the feeding sequence). The DUI data for this batch is then fed back to steps S2 and S3 to update the sample library for subsequent model training.

[0042] In step S4.1, to eliminate dimensional differences and unify the data range, the dynamic flow value DFV output from S2 and the power fluctuation variance PFV extracted from S3 are standardized. The standardization formula for DFV is as follows: ; The PFV standardization formula is: ; In the formula: DFV represents the dynamic flow value output by S2 (original range: 0~200mm, corresponding to UHPC slump flow value). DFV min / DFV max This represents the minimum / maximum DFV value in the historical samples (determined through statistical analysis of 100 sets of experiments, e.g., DFV). min =50mm, DFV max =180mm); PFV represents the power fluctuation variance calculated by S3 (original unit: W). 2 Range: 0.5~15 W 2 ); PFV min / PFV max This represents the minimum / maximum PFV value in historical samples (e.g., PFV). min =0.8 W 2 PFVmax =12W 2 ); DFV norm / PFV norm This represents the standardized feature value, with a uniform range of [0,1]. Higher values ​​indicate better liquidity and stronger dispersion stability, respectively.

[0043] In step S4.2, the DFV is adaptively adjusted based on the stirring process, with the time t after the feeding is completed as the variable. norm With PFV norm The weighting is as follows: In the initial stirring stage (t≤30s), material flowability dominates the dispersion process; in the later stage (t>30s), power stability better reflects the fiber dispersion state. The weighting function formula is as follows: ; ; In the formula: t represents the current mixing time (unit: s, starting from the moment the material is fed in). t0 represents the time of the weight inflection point (the experiment determined it to be 30s, meaning that after 30s, the stability weight exceeds the liquidity weight). k represents the weight adjustment rate coefficient (ranging from 0.1 to 0.3, with k=0.2 resulting in a smoother weight transition). w1(t) / w2(t) represents DFV norm / PFV norm The dynamic weights transition smoothly from 1 to 0 and from 0 to 1 as t progresses (w1=1, w2=0 when t=0; w1=0, w2=1 when t→∞).

[0044] In step S4.3, standardized features and dynamic weights are integrated to construct a comprehensive evaluation index, as shown in the following formula: In the formula: Δ(t) represents the stage compensation coefficient (which is linearly adjusted with t, Δ=0.1 when t=0 and Δ=0.05 when t=60s), used to offset the influence of data noise in the initial stage of stirring; DUI(t) represents the real-time dispersion uniformity index (range: 0~1.2), and the higher the value, the better the dispersion (the historical average DUI value for qualified samples is 0.85±0.1).

[0045] From the above, we can conclude that: This invention significantly improves the scenario adaptability and evaluation accuracy of the entire monitoring method by adaptively adjusting the contribution weights of the video dynamic flow value and power fluctuation variance throughout the entire stirring cycle. The algorithm innovatively dynamically allocates feature weights according to the stirring process. In the initial material mixing stage, higher weights are given to the dynamic flow value, which reflects macroscopic fluidity, to capture the flow characteristics of the initial fusion of fibers and matrix. As stirring enters the stable period, it automatically transitions to an evaluation logic dominated by power fluctuation variance, accurately identifying the microscopic stable state after fiber dispersion. This dynamic adjustment mechanism effectively solves the evaluation deviation problem of traditional fixed-weight fusion at different stages of stirring. Furthermore, by standardizing features to unify dimensional differences, combined with the design of stirring stage compensation coefficients and dynamic thresholds, it further enhances the evaluation robustness under complex working conditions. In practical applications, it improves the real-time calculation accuracy of the fiber dispersion uniformity index, enabling more sensitive identification of early fiber agglomeration risks and providing a reliable basis for timely adjustment of stirring parameters.

[0046] In this invention, by acquiring video streams and motor power data in real time during the mixing process, dynamic monitoring of fiber dispersion can be achieved simultaneously with material mixing. When the system detects abnormal fiber dispersion, it can immediately trigger an early warning, allowing operators to adjust mixing parameters on-site, such as extending the mixing time or optimizing the feeding sequence, thus avoiding the scrapping of the entire batch of materials and significantly improving the controllability and efficiency of the production process.

[0047] By fusing visual and electrical data, the macroscopic flow characteristics reflected by video data and the microscopic interaction characteristics reflected by power data complement each other. When the material flow is uniform and the power fluctuation is stable, good fiber dispersion can be cross-validated. If the flow is abnormal but the power is stable, or vice versa, the system can investigate the root cause of the problem through data correlation. This multi-dimensional verification mechanism greatly reduces the probability of misjudgment by a single indicator, making the evaluation results closer to the actual fiber dispersion state.

[0048] In this invention, after training with a large number of labeled samples, the model can automatically extract flow features from videos, such as vortex morphology and material diffusion speed, and output quantitative flow indicators without human intervention. Simultaneously, power data fluctuation analysis is also automatically completed through the algorithm. The entire monitoring process, from data acquisition to result output, is fully automated, avoiding human error and reducing reliance on highly skilled workers, thus enabling the production line to maintain consistent monitoring accuracy.

[0049] In this invention, non-contact monitoring is achieved through an industrial camera and a power sensor. The camera is installed outside the mixing tank, and power data is collected indirectly through the sensor. The entire process does not involve physical contact with the material, does not change the original state of the material, and does not affect the normal operation of the mixing equipment. This non-invasive design allows monitoring to be carried out synchronously with the production process without interrupting production or sacrificing test specimens, significantly improving the feasibility of industrial applications.

[0050] In this invention, when applied to glass fiber reinforced cement, only the flowability characteristic parameters in the video model need to be adjusted to adapt to the flow characteristics of the cement, or the power fluctuation threshold needs to be modified to match the interaction strength between the glass fiber and the matrix. For carbon fiber reinforced plastics, the camera's spectral range can be optimized to meet the imaging requirements of dark materials. This flexibility allows the method to be reused across material types, providing a universal solution for the mixing quality control of different fiber-reinforced composite materials, thus expanding the application scenarios and industrial value of the technology.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring concrete mixing and evaluating fiber dispersibility based on data fusion, characterized in that: Includes the following steps S1: Fix an industrial camera in the observation window or on top of the mixer to record the material flow video of the mixing process at a fixed frame rate; synchronously collect real-time power data of the motor through a power sensor with a sampling frequency ≥10Hz; perform anti-shake and noise reduction preprocessing on the video data, and perform outlier removal and smoothing on the power data to ensure that the data quality meets the requirements of the subsequent flowability assessment in S2 and stability analysis in S3. S2: Input the video stream preprocessed in S1 into the trained flowability assessment model. This model is trained based on historical mixing video samples and corresponding slump flow value labels, and can output dynamic flow values ​​that reflect the macroscopic flow state of the material. Calculate the dynamic flow values ​​of the current 30-second video window in real time. The results will be used together with the power fluctuation characteristics output in S3 as the fusion basis for S4. S3: Extract the time series of the stable stage in the later stage of stirring from the power data collected in S1, and calculate the power fluctuation variance during this period; experimental verification shows that the variance value is negatively correlated with the fiber dispersion uniformity. This characteristic value will be input into S4 together with the dynamic flow value of S2 for comprehensive evaluation. S4: Based on the dynamic flow value output by S2 and the power fluctuation variance extracted by S3, a dynamic weighted fusion model is constructed; the weights of the two indicators are automatically adjusted according to the stirring process, and the fiber dispersion uniformity index is obtained through weighted calculation; a threshold is set in combination with historical qualified sample data, and an alarm is triggered when the index is lower than the threshold, so as to realize the real-time judgment of stirring quality.

2. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S1, video data acquisition requires a fixed industrial camera with a resolution of no less than 1920×1080 at the observation window or top center of the mixer. The lens focal length is adjusted according to the diameter of the mixing drum to ensure complete coverage of the material flow area. The frame rate is set to 25 frames per second, and the recording period is from the start of feeding until 10 seconds after the end of mixing. The focus is on capturing the dynamic process of material tumbling and flowing after feeding. Synchronous power data acquisition is initiated using a current-voltage combined power sensor with an accuracy class of 0.

5. The motor power signal is acquired in real time through the inverter's analog output interface, and the sampling frequency needs to reach 10 Hz or higher to ensure that the power fluctuation details at the 0.1-second level during the fiber dispersion process can be captured.

3. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S1, during the data preprocessing stage, the video data is first enhanced with grayscale to eliminate uneven brightness caused by flickering stirring lights. Then, Gaussian filtering is used to remove dust noise from the industrial environment. Finally, optical flow is used to register adjacent frames to eliminate image shift caused by slight camera vibration. The power data preprocessing uses the 3σ criterion to remove peak outliers at the moment of motor startup. Then, a moving average filter with a sliding window size of 50 sampling points is used for smoothing to ensure that the data fluctuation characteristics are not affected by high-frequency noise. The processed video and power data are stored as timestamp-aligned data streams to provide synchronous input for the flow assessment in S2 and the stability analysis in S3.

4. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S2, the model training phase requires the construction of a sample library containing more than 200 sets of different working conditions. Each set of samples covers UHPC stirring videos with fiber volume content of 1% to 3% and water-cement ratio of 0.2 to 0.

3. Simultaneously, the 30-second flow value of the corresponding sample is measured through a collapse spread test as a label. The acquisition of the dataset specifically includes: recording the entire mixing process and monitoring power data simultaneously, then measuring the flowability through a slump test, and after the specimen hardens, cutting it and performing image processing to obtain the corresponding fiber dispersion data; thereby obtaining the most suitable flowability for fiber dispersion, and then using image + power to learn, the image can identify the surface of the slurry to ensure good fiber dispersion flowability, and the power can reflect the internal uniformity during the mixing process. The flow value range is controlled between 150 mm and 250 mm. A hybrid architecture of 3D convolutional neural network and long short-term memory network is used as the flow evaluation model. The 3D CNN layer extracts the spatial flow features in the video clips, and the LSTM layer captures the flow trend changes in the time dimension. The model input is a video clip of 150 consecutive frames, and the output is a dynamic flow value DFV normalized to 0 to 100. The mean squared error loss function is used during training, and the iteration is until the error on the validation set is less than 5 mm.

5. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S2, during online application, the system captures the video stream of the first 30 seconds of the current moment in real time, divides it into 5 consecutive segments using a 6-second sliding window, inputs each segment into the pre-trained model to obtain 5 intermediate DFV values, and calculates the final real-time DFV by weighted average. The weight allocation is dynamically adjusted according to the stirring process, with the weight biased towards the later segments within 10 seconds after material addition to reduce interference in the initial mixing stage. This DFV value will be transmitted to S4 in real time and participate in the calculation of the dispersion uniformity index together with the PFV output from S3.

6. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S3, the power data extraction first requires determining the starting point of the stable phase using a stirring process identification algorithm. The stable phase begins when the absolute value of the power change rate is less than 0.5% for 5 consecutive seconds, typically corresponding to the 20-30 second interval after feeding. The extraction duration is fixed at 20 seconds to ensure data consistency. Feature extraction is then performed on the extracted power time series. First, the average power μ within that time period is calculated, and then the power is calculated using the formula PFV=Σ(pi-μ). 2 / (n-1) calculates the fluctuation variance, where pi is the power value of the i-th sampling point, n is the number of sampling points, and PFV is in watt-squared, reflecting the degree of dispersion of power from the mean per unit time.

7. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 1, characterized in that: In step S4, the specific process includes: S4.1 Data Spatiotemporal Alignment: The dynamic flow value DFV output in step S2 and the power fluctuation variance PFV extracted in step S3 in the same stirring batch are spatiotemporally matched to ensure that they correspond to the same stirring stage; data alignment is achieved through timestamp synchronization to provide consistent input for subsequent fusion calculations; S4.2 Linear weighted fusion calculation: The fiber dispersion uniformity index DUI is constructed using a linear weighting method, as shown in the following formula: DUI = a·DFV−b·PFV+c; in: DFV is the dynamic flow value output in step S2, which reflects the macroscopic flowability of the material. The higher the value, the more uniform the flow state. PFV is the power fluctuation variance calculated in step S3, which reflects the stirring stability. The lower the value, the more uniform the fiber dispersion. a, b, and c are coefficients fitted by regression of experimental data. They were obtained by collecting DFV, PFV, and corresponding fiber dispersion coefficients of hardened sections from more than 100 qualified samples and solving them using the least squares method. S4.3 Dynamic threshold determination and feedback: Set the qualified threshold based on the DUI statistics of historical qualified batches; After calculating the DUI of the current batch in real time, if it is lower than the threshold, the system will automatically trigger an alarm, prompting the operator to adjust the stirring parameters, and feed back the DUI data of the batch to steps S2 and S3 for updating the sample library for subsequent model training.

8. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 7, characterized in that: In step S4.1, to eliminate dimensional differences and unify the data range, the dynamic flow value DFV output in S2 and the power fluctuation variance PFV extracted in S3 are standardized, wherein the DFV standardization formula is: ； The PFV standardization formula is: ； In the formula: DFV represents the dynamic flow value output by S2; DFV min / DFV max This represents the minimum / maximum DFV value in the historical samples; PFV represents the power fluctuation variance calculated by S3; PFV min / PFV max This represents the minimum / maximum PFV value in the historical samples. DFV norm / PFV norm This represents the standardized feature value, with a uniform range of [0,1]. Higher values ​​indicate better liquidity and stronger dispersion stability, respectively.

9. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 7, characterized in that: In step S4.2, the DFV is adaptively adjusted based on the stirring process, with the time t after the feeding is completed as the variable. norm With PFV norm The weighting is as follows: In the initial stirring stage (t≤30s), material flowability dominates the dispersion process; in the later stage (t>30s), power stability better reflects the fiber dispersion state. The weighting function formula is as follows: ； ； In the formula: t represents the current mixing time (unit: s, starting from the moment the material is fed in). t0 represents the time of the weight inflection point (the experiment determined it to be 30s, meaning that after 30s, the stability weight exceeds the liquidity weight). k represents the weight adjustment rate coefficient (ranging from 0.1 to 0.3, with k=0.2 resulting in a smoother weight transition). w1(t) / w2(t) represents DFV norm / PFV norm The dynamic weights transition smoothly from 1 to 0 and from 0 to 1 as t progresses.

10. The method for monitoring concrete mixing and evaluating fiber dispersion based on data fusion as described in claim 7, characterized in that: In step S4.3, standardized features and dynamic weights are integrated to construct a comprehensive evaluation index, as shown in the following formula: In the formula: Δ(t) represents the stage compensation coefficient (which is linearly adjusted with t, Δ=0.1 when t=0 and Δ=0.05 when t=60s), used to offset the influence of data noise in the initial stage of stirring; DUI(t) represents the real-time dispersion uniformity index; a higher value indicates better dispersion.

Images