A precise concrete placement intelligent pouring control system and method
By combining multi-task visual perception and intelligent control modules, the problems of manual dependence, single detection and low degree of automation in existing pouring control technology are solved, and high-precision and stable pouring process control and unmanned production are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CCCC THIRD NAVIGATION (NANTONG) OFFSHORE ENG CO LTD
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-02
Smart Images
Figure CN122131718A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pouring control technology, specifically to an intelligent pouring control system and method for precise concrete placement. Background Technology
[0002] In industrial production, the casting of liquid and semi-fluid materials is a core process in the production of precast concrete components, chemical molded parts, and other products. The uniformity of material distribution, the accuracy of material injection, and the consistency of the process directly determine the molding quality, structural strength, and performance of the finished product. To meet the requirements of industrialized mass production, the casting process must achieve precise and stable material injection. The dynamic adjustment capability of the material flow rate is the core to ensure the uniformity and integrity of the casting, while the consistency between the actual opening and closing state of the material hopper and other actuators and the control commands is a crucial link to ensure the reliability and controllability of the casting process.
[0003] The actual industrial pouring site conditions are complex and variable. Material flow rate fluctuates randomly due to factors such as slump and temperature, and problems such as uneven liquid accumulation and local overflow are prone to occur during the pouring process. At the same time, the material hopper gate is prone to mechanical jamming and structural wear. Interference from dust, strong light, shadow, and vibration in the industrial environment also places stringent requirements on the detection and control of the pouring status and actuators. Against this backdrop, traditional pouring control strategies struggle to balance detection accuracy and control robustness. Existing pouring control methods in the industry are mainly divided into three categories: fixed opening or timed control, single-view visual judgment, and physical sensor feedback of hopper gate status. Although these methods have achieved a certain degree of preliminary automation of the pouring process, they all have obvious technical limitations and cannot adapt to the refined and intelligent control requirements under complex working conditions.
[0004] Currently, mainstream concrete segment casting production lines still largely rely on manual intervention and supplemented by automated detection processes. Although some production lines are equipped with basic detection devices such as weight sensors, monocular cameras, and contact level gauges, the control of core process parameters still depends entirely on the subjective experience of the operators. Workers must remain by the mold throughout the process, visually observing the accumulation pattern and flow rate of the concrete surface, and manually adjusting the opening and closing of the material hopper gate and the working frequency of the vibrator. The information collected by various detection devices is only used for a final rough judgment of the casting completion and does not participate in the dynamic control during the process, resulting in the industry status quo of "manual adjustment of the core and equipment assistance for the finishing touches."
[0005] Based on practical application scenarios and technical implementation logic, existing pouring control technologies have many prominent defects, which have become the core obstacle restricting the automation and unmanned upgrading of the pouring process. Specifically, these defects are reflected in the following aspects:
[0006] The process relies heavily on manual experience and lacks quantitative control standards: key operations such as opening and closing the material hopper gate and adjusting the vibration frequency are all performed by workers based on experience. There are no unified quantitative process indicators, and the operating habits and judgment standards of different workers vary significantly, resulting in poor consistency and low repeatability of the pouring quality. At the same time, manual judgment is easily affected by factors such as fatigue and distraction, which can easily lead to operational errors and fluctuations in product quality.
[0007] Without a real-time feedback mechanism, closed-loop control is difficult to achieve: The data collected by the existing detection devices does not participate in the control of the pouring process, but is only used for the final threshold judgment of "whether the pouring is completed". Workers cannot obtain timely feedback on the execution effect of operation instructions, and the adjustment behavior has obvious lag. When abnormal situations such as rapid rise of liquid level or local accumulation occur, manual intervention is often difficult to respond quickly, which can easily lead to production accidents such as material overflow and under-pouring.
[0008] The detection methods are limited, the information dimensions are limited, and the robustness is insufficient: Existing systems mostly use a single method among weight detection, monocular vision detection, or contact-based liquid level detection as the basis for judgment, which cannot comprehensively characterize the spatial distribution of materials within the mold. In scenarios such as concrete adhering to the walls, local accumulation, liquid level fluctuations, and foam interference, single detection information is prone to deviation from the actual pouring state, resulting in low accuracy in judging the liquid level and poor system adaptability to complex working conditions.
[0009] Visual and physical inspection each have their own technical shortcomings and poor adaptability: monocular vision-based inspection solutions can only acquire two-dimensional image information, lacking the ability to perceive the spatial shape and depth distribution of the liquid surface, and are easily interfered with by factors such as dust, reflection, and mold obstruction in the industrial environment, resulting in blurred liquid surface boundaries and false detections; weight detection judges the completion of casting based solely on the overall weight of the mold, and cannot distinguish anomalies such as uneven material distribution, wall residue, and foreign matter mixing, which easily leads to misjudgments; contact liquid level sensors need to be in direct contact with high-viscosity and highly abrasive materials, which are prone to adhesion, clogging, and wear, resulting in decreased detection accuracy, high failure rate, and high maintenance costs, making it difficult to adapt to the high-frequency and continuous production needs of production lines.
[0010] The detection and control logic is simple and lacks the ability to make comprehensive judgments based on multi-source information: existing liquid level detection and process parameter control mostly adopt threshold comparison and simple rule judgment methods, which lack the ability to comprehensively analyze multi-dimensional information such as changes in liquid level shape, weight change trends, and equipment operating status. It is unable to intelligently distinguish complex pouring states, and problems such as false stops, missed stops, and false alarms are prone to occur in practical applications, which seriously affect production efficiency and product quality.
[0011] Low level of automation, unable to support the demand for unmanned production: The manual intervention mode of core process parameters makes it impossible for the existing pouring system to achieve truly unmanned continuous production, which is contrary to the industrial upgrading direction of "lights-out factories" and has become a key process bottleneck restricting the intelligent and automated transformation of the manufacturing industry.
[0012] In summary, existing industrial fluid material pouring control technologies suffer from high reliance on manual labor, limited detection methods, lack of closed-loop control, and insufficient comprehensive judgment capabilities. These limitations make it difficult to simultaneously achieve accurate perception of liquid level status and reliable verification of actuator actions under complex industrial conditions. The industry urgently needs a pouring control solution capable of simultaneously sensing both "material status" and "actuator status," integrating multi-source information for intelligent decision-making. This solution should utilize a fully closed-loop adaptive control mechanism to improve the detection accuracy, control stability, and intelligence level of the pouring process, thereby meeting the demands for unmanned and precise industrial production. Summary of the Invention
[0013] The purpose of this invention is to provide a precise concrete placement intelligent pouring control system and method to solve the problems in the prior art.
[0014] To achieve the above objectives, the present invention provides the following technical solution: a precision material placement intelligent pouring control system, comprising a multi-task visual perception module, a pouring decision module, and an intelligent control module, wherein each module interacts with data and control signals to form a closed-loop control structure;
[0015] The multi-task visual perception module includes a multi-channel industrial camera, a visual data preprocessing unit, and a multi-task visual recognition unit. It is used to acquire dual-view images of the pouring area and complete preprocessing. At the same time, it performs parallel tasks such as cover plate positioning binary classification recognition, concrete liquid surface five-class pouring stage determination, concrete pouring liquid surface instance segmentation, material hopper gate OBB rotating frame target detection and opening / closing angle calculation. It outputs cover plate positioning status, pouring process status information, liquid surface width and area and its changing trend, and hopper gate opening / closing angle.
[0016] The intelligent control module includes an intelligent vibration module and an intelligent material feeding feedback control module. The intelligent vibration module is based on a one-dimensional conditional generative adversarial network and generates a smooth and continuous vibration frequency sequence with time information, visual state and concrete slump as conditional inputs.
[0017] The intelligent feeding feedback control module dynamically adjusts the control parameters of the feeding actuator by constructing a joint feedback correction strategy based on the pouring stage status and the deviation of the hopper gate opening and closing angle.
[0018] The pouring decision module uses a multilayer perceptron (MLP) as its core decision model. It inputs a pouring state vector consisting of liquid surface width, liquid surface area, and real-time weight information, and outputs decision results such as incomplete pouring, completed pouring, and overflow / abnormal pouring, thereby controlling the start and stop of the pouring process.
[0019] Preferably, the visual data preprocessing unit of the multi-task visual perception module performs the following operations:
[0020] Images were acquired using different cameras to obtain dual-view images:
[0021] ;
[0022] Where H and W represent the height and width of the image, respectively, and C represents the number of image channels;
[0023] By performing region recognition on the image, the location boundary (ROI) of the pouring area in the image is determined, and based on the boundary, the region of interest is extracted from the original image to obtain the pouring area image compared to the original image. , Extract sub-images containing the effective pouring area respectively ;
[0024]
[0025]
[0026] in: The coordinates of the upper left corner of the pouring area; These are the area width and height, respectively;
[0027] The region of interest (ROI) image is then defined as:
[0028]
[0029]
[0030] The image of the region of interest is normalized and scaled to a preset size. The processed image is obtained. ;
[0031]
[0032]
[0033] The scaling factor is:
[0034] , , .
[0035] Preferably, the five stages of concrete pouring include: no material piled up, small material piled up, near overflow, not fully poured, and fully poured.
[0036] The multi-task visual recognition unit uses a five-class classification network based on the pouring stage status. Feature extraction and fusion analysis are performed on the preprocessed dual-view images to output the state probabilities of five types of pouring stages, satisfying:
[0037] , ( ),
[0038] The stage corresponding to the maximum probability is taken as the current pouring process state.
[0039] Preferably, the multi-task visual recognition unit uses a liquid surface instance segmentation network. Pixel-level segmentation is performed on the preprocessed dual-view images to obtain liquid surface instance masks. and The segmentation process can be represented as:
[0040] ( ),
[0041] The width of the liquid bottom is calculated using a frontal view mask M1. This is achieved by extracting the edge point set, constructing a flatness function to filter the adaptive effective height range, and then using a random consistency sampling algorithm to fit the edge line and calculate the width value. A side view mask is also used. Calculate the liquid surface area to satisfy:
[0042]
[0043] in, Represents pixels Whether it belongs to the concrete liquid surface area, when =1 indicates that the pixel belongs to the liquid surface area; and These represent the height and width of the mask image, respectively.
[0044] Preferably, the multi-task visual recognition unit outputs the orientation bounding box parameters of the hopper gate through OBB rotating frame target detection:
[0045]
[0046] in: The coordinates of the upper left corner of the hopper gate; These are the width and height of the hopper gate, respectively; This indicates the rotation angle of the bounding box relative to the image coordinate system;
[0047] Based on the geometric principal orientation characteristics of the hopper gate and the relationship between the length and short sides of the enclosing frame, the angle is corrected to obtain the actual orientation angle of the hopper gate:
[0048]
[0049] The angle is periodically normalized to output the final attitude angle of the hopper gate. As an important input parameter for determining the state of the hopper gate and subsequently controlling its operation:
[0050] .
[0051] Preferably, the one-dimensional conditional generative adversarial network of the intelligent vibration module includes a generator. and discriminator ;
[0052] The generator uses a random noise vector and condition vector As a combined input, generate a vibration frequency sequence: ,in, For normalized time variables, This is a value representing the slump of cement.
[0053] The discriminator takes the vibration frequency sequence f and the condition vector c as input, judges the consistency between the sequence authenticity and the condition, and outputs:
[0054]
[0055] The objective function for network training is:
[0056]
[0057] in To counteract the loss function, the second term is a one-dimensional total variation regularization term, where λ is the weighting coefficient and T is the time series length.
[0058] Preferably, the intelligent feeding feedback control module calculates the target adjustment angle of the hopper gate. Compared with the actual angle of visual inspection Attitude deviation is expressed as:
[0059]
[0060] Based on the state s and posture deviation during the pouring stage Construct a material feeding correction mapping function to control the output to meet the requirements. During the initial pouring stage, a large range of attitude adjustment is allowed. As the pouring nears completion, the tolerance for attitude deviation is reduced. In the event of overflow risk, a strategy to slow down or stop material feeding is triggered.
[0061] Preferably, the input to the MLP decision network is the pouring state features output by the multi-task visual perception module and the weight acquisition module, including the width of the bottom of the concrete liquid surface. Liquid surface area And concrete weight information collected in real time by the weighing device. Together, they constitute the pouring state vector at the current moment:
[0062]
[0063] The MLP network extracts and fuses features from the input state vector through a multi-layer fully connected structure. Its decision output is used to represent the control decision result of the current pouring stage. The intelligent decision-making process can be represented as follows:
[0064]
[0065] in, The decision output vector represents three states: incomplete pouring, completed pouring, and overflow or abnormal.
[0066] A precise material placement and intelligent pouring control method, using the aforementioned precise material placement and intelligent pouring control system, includes the following steps:
[0067] S1. System Startup and Parameter Initialization: Initialize the pouring task parameters, visual acquisition equipment parameters, deep learning model parameters, vibration and material feeding control parameters, control thresholds and safety constraint parameters. After completion, enter the pouring monitoring state.
[0068] S2. Multi-task visual perception and status information acquisition: The multi-task visual perception module collects images of the pouring area, completes preprocessing and multi-task recognition, and extracts the status of the cover plate in place, the position and height of the concrete liquid surface, and the material placement status of the pouring area. When the cover plate is in place, the subsequent process is entered; if it is not in place, pouring is prohibited or delayed.
[0069] S3. Pouring status judgment: Based on the visual perception results, determine whether it is a valid pouring status, and further determine the current pouring process stage, including the initial pouring stage, the intermediate filling stage, and the near-complete stage. If the pouring conditions are not met, continue to detect; if they are met, proceed to the next step.
[0070] S4. Material feeding and vibration demand judgment: After confirming the effective pouring state, judge whether it is necessary to continue feeding, adjust the feeding speed / amount, and adjust the vibration frequency / strategy based on visual feedback information.
[0071] S5. Pouring Decision Generation: Based on the current pouring process status, time information within the stage, and weight sensor data, a pouring decision is generated through the MLP decision network to determine whether the pouring is complete.
[0072] S6. Safety verification and modulation of control parameters: The generated feeding parameters and vibration control parameters are subjected to amplitude limiting, smoothing filtering and abnormal parameter removal to ensure that the equipment operation safety and process requirements are met.
[0073] S7. Execution control and material feeding and vibration control: Send the verified control parameters to the material feeding device and vibration actuator to execute the material feeding and vibration operations;
[0074] S8. Status Feedback and Closed-Loop Correction: Continuously obtain the latest pouring status through the multi-task visual perception module and feed it back to the pouring decision module to dynamically correct the material feeding and vibration parameters to prevent over-vibration, under-vibration or uneven material distribution.
[0075] S9. Pouring Completion Judgment and Process End: When the pouring decision module determines that the pouring has reached the preset completion conditions, it stops the material feeding and vibration operation, outputs the pouring completion signal, and ends the current pouring process.
[0076] Preferably, the initial pouring stage in step S3 is a state without material pile, the intermediate filling stage is a state of small material pile / large material pile, and the near-complete stage is a state of not being fully poured / fully poured; in step S8, the opening and closing angle of the material hopper gate and the frequency of the vibrator are dynamically closed-loop corrected by real-time collected information on the opening and closing angle of the material hopper gate and the liquid level, so as to realize adaptive control of the pouring process.
[0077] Compared with the prior art, the beneficial effects of the present invention are:
[0078] 1. Enhanced Liquid Surface Sensing Accuracy and State Recognition Capability: This invention utilizes a dual-view industrial camera to determine five types of pouring stages and segment liquid surfaces at the pixel level. This overcomes the insufficient accuracy of traditional monocular vision or weight sensors, accurately sensing the spatial distribution and local accumulation state of the liquid surface, enabling high-precision, fine-grained monitoring of the pouring process. It reduces the risk of under-pouring or overflow due to errors in liquid surface judgment, improving pouring uniformity and product quality consistency.
[0079] 2. Achieving Closed-Loop Feedback Control of Hopper Gate Status: This invention uses an OBB rotating frame to detect and identify the posture and opening / closing angle of the hopper gate in real time, and combines this with the status of the pouring stage for intelligent feeding feedback control, solving the problem of difficulty in real-time verification of the hopper gate's opening and closing status in existing technologies. This enables adaptive correction during the feeding process, ensuring uniform material coverage, reducing local accumulation or overflow, and improving the stability of the pouring process.
[0080] 3. Adaptive Optimization of Vibration Process: This invention generates time-series vibration frequencies based on a one-dimensional conditional generative adversarial network, and combines these with the pouring process conditions, time, and slump conditions to achieve dynamic adjustment and optimization of vibration parameters. This ensures continuous and smooth vibration, adapts to changes in concrete properties, reduces reliance on manual intervention, and improves concrete density and molding quality.
[0081] 4. Achieving Closed-Loop Intelligent Control and Unmanned Production: Through the combined application of multi-task visual perception, hopper gate posture feedback, and vibration adaptive control, this invention constructs a fully closed-loop control mechanism of "perception → decision-making → execution → verification." This significantly improves the adaptability, safety, and intelligence level of the pouring process, supports unattended continuous production, and meets the construction needs of "lights-out" factories. Attached Figure Description
[0082] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0083] Figure 1 This is a schematic diagram of the structure of the present invention;
[0084] Figure 2 This is a mapping block diagram of the intelligent decision-making module of the present invention;
[0085] Figure 3 This is a flowchart of the present invention. Detailed Implementation
[0086] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.
[0087] Please see Figure 1-3 In this embodiment of the invention, a precise material placement intelligent pouring control system includes a multi-task visual perception module, a pouring decision module, and an intelligent control module. Each module interacts with data and control signals to form a closed-loop control structure.
[0088] The multi-task visual perception module includes a multi-channel industrial camera, a visual data preprocessing unit, and a multi-task visual recognition unit. It is used to acquire dual-view images of the pouring area and complete preprocessing. At the same time, it performs parallel tasks such as cover plate positioning binary classification recognition, concrete liquid surface five-class pouring stage determination, concrete pouring liquid surface instance segmentation, material hopper gate OBB rotating frame target detection and opening / closing angle calculation. It outputs cover plate positioning status, pouring process status information, liquid surface width and area and its changing trend, and hopper gate opening / closing angle.
[0089] The intelligent control module includes an intelligent vibration module and an intelligent material feeding feedback control module. The intelligent vibration module is based on a one-dimensional conditional generative adversarial network and generates a smooth and continuous vibration frequency sequence with time information, visual state and concrete slump as conditional inputs.
[0090] The intelligent feeding feedback control module dynamically adjusts the control parameters of the feeding actuator by constructing a joint feedback correction strategy based on the pouring stage status and the deviation of the hopper gate opening and closing angle.
[0091] The pouring decision module uses a multilayer perceptron (MLP) as its core decision model. It inputs a pouring state vector consisting of liquid surface width, liquid surface area, and real-time weight information, and outputs decision results such as incomplete pouring, completed pouring, and overflow / abnormal pouring, thereby controlling the start and stop of the pouring process.
[0092] Processing flow
[0093] The first step is to acquire images from different cameras to obtain dual-view images:
[0094]
[0095] Where H and W represent the height and width of the image, respectively, and C represents the number of image channels.
[0096] Step 2: By performing region recognition on the image, the location boundary (ROI) of the pouring area in the image is determined, and based on the boundary, the region of interest is extracted from the original image to obtain the pouring area image compared to the original image. , Extract sub-images containing the effective pouring area respectively ;
[0097] Let the bounding boxes of the pouring areas in the image be represented as follows:
[0098]
[0099]
[0100] in: The coordinates of the upper left corner of the pouring area; These represent the area width and height, respectively.
[0101] The region of interest (ROI) image is then defined as:
[0102]
[0103]
[0104] To meet the input requirements of subsequent visual analysis and decision-making models, the image of the region of interest is normalized and scaled to a preset size. The processed image is obtained. :
[0105]
[0106]
[0107] The scaling factor is: , ,
[0108] Step 3: Recognize specific images input for different tasks:
[0109] 1. For the original image from a single viewpoint Perform cover plate positioning identification:
[0110] A single-view original image cover plate positioning recognition method based on a classification network is adopted. The single-view original image... After preprocessing such as size normalization and pixel normalization, the result is obtained .
[0111]
[0112] This represents the image preprocessing mapping function.
[0113] The cover plate positioning recognition adopts a binary classification method based on a classification network. The classification network uses the normalized image... As input, the output corresponds to the state probability value and confidence level of "cover in place" and "cover not in place", and the determination process can be expressed as follows:
[0114] , ( ), + =1
[0115] in, A classification network representing the position status of the cover plate. This indicates the confidence level of whether the cover plate is in place or not.
[0116] When satisfied If the cover plate is in the correct position, output classification result 1; otherwise, if the cover plate is not in the correct position, output 0. To pre-set the reliability threshold.
[0117] 2. Identification and determination of the state of concrete liquid surface during the pouring stage from a dual-view perspective;
[0118] In this invention, a five-class classification method based on dual-view images is employed to determine the phased changes in the concrete surface pouring process. The system simultaneously acquires dual-view raw images of the pouring area using two industrial cameras positioned at different locations. and After preprocessing such as size normalization, the data are input into the pouring stage state recognition and classification network for feature extraction and fusion analysis, outputting the state probability and confidence level corresponding to different pouring stages. The determination process can be expressed as follows:
[0119] , ( ),
[0120] in, A five-class network representing the state of the pouring stage. Indicates that the image belongs to the first... Confidence level for the pouring stage. When the following conditions are met:
[0121]
[0122] When the system determines the current pouring state as the s-th pouring stage, it uses this stage determination result as the state input for subsequent pouring decisions and material feeding control. The five pouring stage states include at least: no material piled up, small material piled up, adjacent overflow, not fully poured, and fully poured.
[0123] 3. Perform instance segmentation of the concrete pouring liquid surface from both perspectives;
[0124] In this invention, to achieve refined perception of the spatial distribution and local morphology of the concrete pouring surface, a concrete surface instance segmentation method based on dual-view images is employed. The pre-processed... The data are input into the liquid surface instance segmentation network to perform pixel-level segmentation of the concrete liquid surface region, thereby obtaining the liquid surface instance mask from the corresponding viewpoint. and The segmentation process can be represented as:
[0125] ( ),
[0126] in, Represents the concrete liquid surface instance segmentation network. The results of liquid surface instance segmentation under the corresponding viewpoint are used to characterize the pixel-level distribution area of concrete liquid surface in the image.
[0127] Further, different characterization values are obtained for the instance segmentation results output from two different perspectives. The first perspective, the frontal view, allows for a relatively clear observation of the liquid surface, but some hopper doors obstruct the view, so it is used to calculate the width of the bottom of the liquid surface. The second perspective, the side view, allows for a clear observation of the entire liquid surface, and is used to calculate the area A of the liquid surface.
[0128] Calculation of bottom width of liquid surface:
[0129] Extracting the liquid surface edge point set from the liquid surface mask and in vertical height A geometric flatness function is constructed based on the lateral geometric distribution of edge points:
[0130]
[0131] Simultaneously, a confidence flatness function is constructed based on the segmentation confidence corresponding to the edge points:
[0132]
[0133] Based on this, a liquid surface edge smoothness function is constructed by combining geometric undulations and confidence stability:
[0134]
[0135] Based on the edge smoothness function, longitudinal height positions that meet the stability conditions are dynamically selected to determine the adaptive effective height range for liquid surface geometry calculations:
[0136]
[0137] Select the effective set of points at the edge of the liquid surface within the dynamic height range:
[0138]
[0139] Robust line fitting is performed on the effective edge point set using a random consistency sampling algorithm to obtain two straight lines in the main direction model of the liquid surface edge. , .
[0140]
[0141]
[0142] Based on the boundary position of the fitted model within the dynamic height range, the unfolded width of the bottom of the concrete liquid surface is calculated. :
[0143]
[0144] liquid surface area calculate:
[0145]
[0146] in, Represents pixels Whether it belongs to the concrete liquid surface area, when =1 indicates that the pixel belongs to the liquid surface area; and These represent the height and width of the mask image, respectively.
[0147] 4. Target detection and recognition of the OBB rotating frame of a single-view hopper gate, and calculation of the opening and closing angle.
[0148] This invention employs an oriented bounding box (OBB)-based target detection method to perform attitude recognition on a hopper gate. The detection network outputs the oriented bounding box parameters of the hopper gate.
[0149]
[0150] in: The coordinates of the upper left corner of the hopper gate; These are the width and height of the hopper gate, respectively; This represents the rotation angle of the bounding box relative to the image coordinate system. Based on the geometric principal orientation characteristics of the hopper gate and the relationship between the length and short sides of the bounding box, the angle is corrected to obtain the actual orientation angle of the hopper gate.
[0151]
[0152] The angle is then periodically normalized to output the final attitude angle of the hopper gate. It serves as an important input parameter for determining the state of the hopper gate and for subsequently controlling its operation.
[0153]
[0154] The aforementioned visual tasks can be executed in parallel, enabling multi-task perception of the pouring process.
[0155] 5. Output
[0156] The output of the multi-task visual perception module includes:
[0157] Cover plate in position p;
[0158] Current pouring process status information ;
[0159] Information on the width W and area A of the liquid surface, as well as their changing trends;
[0160] Hopper gate opening angle ;
[0161] The above outputs serve as the core input data for the intelligent control module and the pouring decision module.
[0162] III. Intelligent Control Module
[0163] 1. Intelligent Vibration Module
[0164] In this invention, the intelligent control module adaptively generates vibration control parameters based on the characteristics of concrete properties changing over time during the pouring process, employing a vibration frequency generation method based on a one-dimensional conditional generative adversarial network. The control module uses time information... and current pouring process status information s, cement slump As conditional input, construct the conditional vector:
[0165]
[0166] in, The normalized time variable is s, which represents the current pouring process status information. This refers to the slump value of cement. The cement slump data is obtained from the mixing plant system of the previous process before material feeding and transmitted to the intelligent control module of this invention via an industrial communication interface, serving as a priori process conditions for vibration control.
[0167] The generator employs a one-dimensional generative network structure, using random noise vectors. With the condition vector As a joint input, a vibration frequency sequence for the corresponding time period is generated:
[0168]
[0169] in,
[0170]
[0171] This represents the vibration frequency output in one-dimensional time series form.
[0172] The discriminator employs a one-dimensional discriminant network structure, taking a real or generated vibration frequency sequence and its corresponding condition vector as input. It judges the authenticity of the vibration frequency sequence in the time dimension and its consistency with the condition information. Its output is expressed as follows:
[0173]
[0174] During network training, the generated vibration frequency sequence is jointly optimized by using an adversarial loss function and introducing a smoothing constraint term for the one-dimensional time series. The objective function is expressed as:
[0175]
[0176] The second term is a one-dimensional total variation regularization term, which is used to suppress high-frequency abrupt changes in the vibration frequency in the time dimension, ensuring the continuity and smoothness of the generated control commands and the executability of the vibration equipment.
[0177] The generated vibration frequency sequence, after being processed by safety limiting, is input as the control setpoint into the vibration actuator to realize adaptive vibration control based on a one-dimensional conditional generation network.
[0178] 2. Intelligent feeding feedback control
[0179] In this invention, the material feeding process employs a joint feedback correction control strategy based on the pouring stage status and hopper gate posture to adapt to the differentiated requirements for feeding accuracy and material distribution uniformity at different pouring stages. The intelligent control module uses the pouring stage status and hopper gate posture information output by the multi-task visual perception module as the main feedback basis to adjust and control the material feeding process online.
[0180] The pouring stage s state is obtained by the visual perception module and is used to characterize the current process stage of the pouring; the hopper gate opening angle... The target adjustment angle is obtained from the target detection and recognition of the OBB rotating frame of the hopper gate from a single perspective and the calculation of the opening and closing angle. Compared with the actual angle of visual inspection Attitude deviation is expressed as:
[0181]
[0182] The intelligent material feeding feedback control module dynamically corrects the control parameters of the material feeding actuator based on the posture deviation and the current pouring stage status. Its control output is expressed as follows:
[0183]
[0184] Where s represents the current pouring stage state, and G(⋅) represents the material feeding correction mapping function constructed based on rules or learning models.
[0185] Specifically, during the initial pouring stage, the system allows for a large range of posture adjustment to ensure material coverage; as the pouring nears completion, the system's tolerance for posture deviations decreases to avoid local accumulation or overflow; when in an abnormal or overflow-risk state, the system can trigger a material feeding slowdown or pause strategy.
[0186] By jointly modeling the pouring stage status with OBB attitude feedback, the material feeding control process can dynamically adjust the control strategy according to the process stage, realize adaptive correction of the hopper gate opening and closing angle, and significantly improve the stability and material distribution accuracy of the material feeding process.
[0187] IV. Intelligent Decision-Making Module
[0188] The intelligent decision-making module is used to comprehensively analyze multi-source state information during the concrete pouring process and generate decision instructions for material feeding and vibration control based on a data-driven approach. The intelligent decision-making module uses a multi-layer perceptron (MLP) network as its core decision model to achieve a nonlinear mapping from the pouring process state to control decisions.
[0189] The input to the MLP decision network is the pouring state features output by the multi-task visual perception module and the weight acquisition module, including the width of the bottom of the concrete liquid surface. Liquid surface area And concrete weight information collected in real time by the weighing device. The above parameters together constitute the pouring state vector at the current moment:
[0190]
[0191] The MLP network extracts and fuses features from the input state vector through a multi-layer fully connected structure. Its decision output is used to represent the control decision result of the current pouring stage, and is expressed as follows:
[0192] The intelligent decision-making process can be represented as follows:
[0193]
[0194] in, The decision output vector represents three states: incomplete pouring, completed pouring, and overflow or abnormal. It is directly used to control the end of the pouring process.
[0195] By introducing based The intelligent decision-making module of this invention can fully explore the nonlinear relationship between the geometric features of the liquid surface and the weight information, realize adaptive decision-making for the pouring process, significantly reduce reliance on human experience, and improve the stability, consistency and intelligence level of the concrete pouring process.
[0196] A precise material placement and intelligent pouring control method, using the aforementioned precise material placement and intelligent pouring control system, includes the following steps:
[0197] S1. System Startup and Parameter Initialization: Initialize the pouring task parameters, visual acquisition equipment parameters, deep learning model parameters, vibration and material feeding control parameters, control thresholds and safety constraint parameters. After completion, enter the pouring monitoring state.
[0198] S2. Multi-task visual perception and status information acquisition: The multi-task visual perception module collects images of the pouring area, completes preprocessing and multi-task recognition, and extracts the status of the cover plate in place, the position and height of the concrete liquid surface, and the material placement status of the pouring area. When the cover plate is in place, the subsequent process is entered; if it is not in place, pouring is prohibited or delayed.
[0199] S3. Pouring status judgment: Based on the visual perception results, determine whether it is a valid pouring status, and further determine the current pouring process stage, including the initial pouring stage, the intermediate filling stage, and the near-complete stage. If the pouring conditions are not met, continue to detect; if they are met, proceed to the next step.
[0200] S4. Material feeding and vibration demand judgment: After confirming the effective pouring state, judge whether it is necessary to continue feeding, adjust the feeding speed / amount, and adjust the vibration frequency / strategy based on visual feedback information.
[0201] S5. Pouring Decision Generation: Based on the current pouring process status, time information within the stage, and weight sensor data, a pouring decision is generated through the MLP decision network to determine whether the pouring is complete.
[0202] S6. Safety verification and modulation of control parameters: The generated feeding parameters and vibration control parameters are subjected to amplitude limiting, smoothing filtering and abnormal parameter removal to ensure that the equipment operation safety and process requirements are met.
[0203] S7. Execution control and material feeding and vibration control: Send the verified control parameters to the material feeding device and vibration actuator to execute the material feeding and vibration operations;
[0204] S8. Status Feedback and Closed-Loop Correction: Continuously obtain the latest pouring status through the multi-task visual perception module and feed it back to the pouring decision module to dynamically correct the material feeding and vibration parameters to prevent over-vibration, under-vibration or uneven material distribution.
[0205] S9. Pouring Completion Judgment and Process End: When the pouring decision module determines that the pouring has reached the preset completion conditions, it stops the material feeding and vibration operation, outputs the pouring completion signal, and ends the current pouring process.
[0206] In step S3, the initial pouring stage is in the state of no material pile, the intermediate filling stage is in the state of small material pile / large material pile, and the near-complete stage is in the state of not fully poured / fully poured. In step S8, the opening and closing angle of the material hopper gate and the frequency of the vibrator are dynamically closed-loop corrected by real-time collected information on the opening and closing angle of the material hopper gate and the liquid level, so as to realize the adaptive control of the pouring process.
[0207] Step S1: System startup and parameter initialization
[0208] After the system starts, the relevant parameters for pouring control are initialized, including:
[0209] Initialize the parameters for the pouring task; initialize the parameters for the visual acquisition equipment and the deep learning model.
[0210] Vibration and feeding control parameters; control thresholds and safety constraint parameters.
[0211] After initialization is complete, the system enters the pouring monitoring state.
[0212] Step S2: Multi-task visual perception and state information acquisition
[0213] The multi-task visual perception module is used to monitor the pouring area in real time, collect image data during the concrete pouring process, and extract the following key information:
[0214] Whether the cover plate is in place determines the start of pouring; when the cover plate is determined to be in place, the system allows the subsequent pouring judgment and material control process to begin; when the cover plate is not in place, the pouring operation is automatically prohibited or delayed, thereby effectively reducing reliance on manual judgment and improving the safety and automation level of the pouring operation.
[0215] Information on the position and height of the concrete liquid level;
[0216] The current material placement status of the pouring area.
[0217] Step S3: Judging the pouring status
[0218] Based on the visual perception results, the current pouring status is judged to determine whether it is in an effective pouring state, and the current process stage of the pouring is further determined.
[0219] The casting process stage includes at least:
[0220] Initial pouring stage: No material piled up; Intermediate filling stage: Small / large material piled up; Near completion stage: Not fully poured / Fully poured (roughly).
[0221] If the current state does not meet the conditions for full pouring, the continuous frame stable state judgment process will be entered and the judgment state will be continuously output.
[0222] If the conditions for full pouring are met, proceed to the next decision-making step.
[0223] Step S4: Material Feeding and Vibration Demand Judgment After confirming the current pouring status is valid, the system determines whether material feeding adjustments or vibration parameter updates are needed based on visual feedback information, including:
[0224] Should we continue feeding materials?
[0225] Do we need to adjust the feeding speed or feeding amount?
[0226] Does it need to be adjusted in terms of vibration frequency or vibration strategy?
[0227] Step S5: Casting Decision Generation
[0228] When parameter adjustments are required, the pouring decision module generates control decisions based on the following information:
[0229] Current pouring process status;
[0230] Time information within the phase;
[0231] Weight sensor data.
[0232] The intelligent decision-making model can directly determine whether the pouring is complete.
[0233] Step S6: Control parameter security verification and modulation
[0234] The generated feeding parameters and vibration control parameters are subjected to safety verification and modulation processing, including:
[0235] Parameter limiting processing;
[0236] Smoothing and filtering processes;
[0237] Abnormal parameters are removed.
[0238] Ensure that the generated control parameters meet the requirements for equipment operation safety and process control.
[0239] Step S7: Execution control and material feeding vibration control
[0240] The verified control parameters are sent to the actuator, including:
[0241] Feeding device;
[0242] Vibration actuator.
[0243] Perform the material feeding and vibration operations according to the control instructions.
[0244] Step S8: State Feedback and Closed-Loop Correction
[0245] During the control process, the system continuously acquires the latest pouring status through the multi-task visual perception module and feeds the status information back to the pouring decision module for:
[0246] Assess the effectiveness of the control measures;
[0247] Dynamically adjust feeding and vibration parameters;
[0248] Prevent over-vibration, under-vibration, or uneven fabric distribution.
[0249] Step S9: Determining completion of pouring and ending the process
[0250] When the intelligent decision-making module determines that the pouring has reached the preset completion conditions, the system stops the material feeding and vibration operations, outputs a pouring completion signal, and ends the current pouring process.
[0251] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A precise material placement intelligent pouring control system, characterized in that, It includes a multi-task visual perception module, a pouring decision module, and an intelligent control module. Each module interacts with data and control signals to form a closed-loop control structure. The multi-task visual perception module includes a multi-channel industrial camera, a visual data preprocessing unit, and a multi-task visual recognition unit. It is used to acquire dual-view images of the pouring area and complete preprocessing. At the same time, it performs parallel tasks such as cover plate positioning binary classification recognition, concrete liquid surface five-class pouring stage determination, concrete pouring liquid surface instance segmentation, material hopper gate OBB rotating frame target detection and opening / closing angle calculation. It outputs cover plate positioning status, pouring process status information, liquid surface width and area and its changing trend, and hopper gate opening / closing angle. The intelligent control module includes an intelligent vibration module and an intelligent material feeding feedback control module. The intelligent vibration module is based on a one-dimensional conditional generative adversarial network and generates a smooth and continuous vibration frequency sequence with time information, visual state and concrete slump as conditional inputs. The intelligent feeding feedback control module dynamically adjusts the control parameters of the feeding actuator by constructing a joint feedback correction strategy based on the pouring stage status and the deviation of the hopper gate opening and closing angle. The pouring decision module uses a multilayer perceptron (MLP) as its core decision model. It inputs a pouring state vector consisting of liquid surface width, liquid surface area, and real-time weight information, and outputs decision results such as incomplete pouring, completed pouring, and overflow / abnormal pouring, thereby controlling the start and stop of the pouring process.
2. The precision material placement intelligent pouring control system according to claim 1, characterized in that, The visual data preprocessing unit of the multi-task visual perception module performs the following operations: Images were acquired using different cameras to obtain dual-view images: ； Where H and W represent the height and width of the image, respectively, and C represents the number of image channels; By performing region recognition on the image, the location boundary (ROI) of the pouring area in the image is determined, and based on the boundary, the region of interest is extracted from the original image to obtain the pouring area image compared to the original image. , Extract sub-images containing the effective pouring area respectively ; in: The coordinates of the upper left corner of the pouring area; These are the area width and height, respectively; The region of interest (ROI) image is then defined as: The image of the region of interest is normalized and scaled to a preset size. The processed image is obtained. ; The scaling factor is: 、 、 。 3. The precision material placement intelligent pouring control system according to claim 1, characterized in that, The five stages of concrete pouring include: no material piled up, small material piled up, near overflow, not fully poured, and fully poured. The multi-task visual recognition unit uses a five-class classification network based on the pouring stage status. Feature extraction and fusion analysis are performed on the preprocessed dual-view images to output the state probabilities of five types of pouring stages, satisfying: , ( ), The stage corresponding to the maximum probability is taken as the current pouring process state.
4. The precision material placement intelligent pouring control system according to claim 3, characterized in that, The multi-task visual recognition unit uses a liquid surface instance segmentation network. Pixel-level segmentation is performed on the preprocessed dual-view images to obtain liquid surface instance masks. and The segmentation process can be represented as: ( )， The width of the liquid bottom is calculated using a frontal view mask M1. This is achieved by extracting the edge point set, constructing a flatness function to filter the adaptive effective height range, and then using a random consistency sampling algorithm to fit the edge line and calculate the width value. A side view mask is also used. Calculate the liquid surface area to satisfy: in, Represents pixels Whether it belongs to the concrete liquid surface area, when =1 indicates that the pixel belongs to the liquid surface area; and These represent the height and width of the mask image, respectively.
5. The precision material placement intelligent pouring control system according to claim 1, characterized in that, The multi-task vision recognition unit outputs the orientation bounding box parameters of the hopper gate through OBB rotating frame target detection: in: The coordinates of the upper left corner of the hopper gate; These are the width and height of the hopper gate, respectively; This indicates the rotation angle of the bounding box relative to the image coordinate system; Based on the geometric principal orientation characteristics of the hopper gate and the relationship between the length and short sides of the enclosing frame, the angle is corrected to obtain the actual orientation angle of the hopper gate: The angle is periodically normalized to output the final attitude angle of the hopper gate. As an important input parameter for determining the state of the hopper gate and subsequently controlling its operation: 。 6. The intelligent pouring control system for precise material placement according to claim 1, characterized in that, The one-dimensional conditional generative adversarial network of the intelligent vibration module includes a generator. and discriminator ; The generator uses a random noise vector and condition vector As a combined input, generate a vibration frequency sequence: ,in, For normalized time variables, This is a value representing the slump of cement. The discriminator takes the vibration frequency sequence f and the condition vector c as input, judges the consistency between the sequence authenticity and the condition, and outputs: The objective function for network training is: in To counteract the loss function, the second term is a one-dimensional total variation regularization term, where λ is the weighting coefficient and T is the time series length.
7. The precision material placement intelligent pouring control system according to claim 1, characterized in that, The intelligent feeding feedback control module calculates the target adjustment angle of the hopper gate. Compared with the actual angle of visual inspection Attitude deviation is expressed as: Based on the state s and posture deviation during the pouring stage Construct a material feeding correction mapping function to control the output to meet the requirements. During the initial pouring stage, a large range of attitude adjustment is allowed. As the pouring nears completion, the tolerance for attitude deviation is reduced. In the event of overflow risk, a strategy to slow down or stop material feeding is triggered.
8. The precision material placement intelligent pouring control system according to claim 1, characterized in that, The input to the MLP decision network is the pouring state features output by the multi-task visual perception module and the weight acquisition module, including the width of the bottom of the concrete liquid surface. Liquid surface area And concrete weight information collected in real time by the weighing device. ; Together they constitute the pouring state vector at the current moment: The MLP network extracts and fuses features from the input state vector through a multi-layer fully connected structure. Its decision output is used to represent the control decision result of the current pouring stage. The intelligent decision-making process can be represented as follows: in, The decision output vector represents three states: incomplete pouring, completed pouring, and overflow or abnormal.
9. A precise intelligent pouring control method for material placement, characterized in that, The precision material placement intelligent pouring control system according to any one of claims 1-8 includes the following steps: S1. System Startup and Parameter Initialization: Initialize the pouring task parameters, visual acquisition equipment parameters, deep learning model parameters, vibration and material feeding control parameters, control thresholds and safety constraint parameters. After completion, enter the pouring monitoring state. S2. Multi-task visual perception and status information acquisition: The multi-task visual perception module collects images of the pouring area, completes preprocessing and multi-task recognition, and extracts the status of the cover plate in place, the position and height of the concrete liquid surface, and the material placement status of the pouring area. When the cover plate is in place, the subsequent process is entered; if it is not in place, pouring is prohibited or delayed. S3. Pouring status judgment: Based on the visual perception results, determine whether it is a valid pouring status, and further determine the current pouring process stage, including the initial pouring stage, the intermediate filling stage, and the near-complete stage. If the pouring conditions are not met, continue to detect; if they are met, proceed to the next step. S4. Material feeding and vibration demand judgment: After confirming the effective pouring state, judge whether it is necessary to continue feeding, adjust the feeding speed / amount, and adjust the vibration frequency / strategy based on visual feedback information. S5. Pouring Decision Generation: Based on the current pouring process status, time information within the stage, and weight sensor data, a pouring decision is generated through the MLP decision network to determine whether the pouring is complete. S6. Safety verification and modulation of control parameters: The generated feeding parameters and vibration control parameters are subjected to amplitude limiting, smoothing filtering and abnormal parameter removal to ensure that the equipment operation safety and process requirements are met. S7. Execution control and material feeding and vibration control: Send the verified control parameters to the material feeding device and vibration actuator to execute the material feeding and vibration operations; S8. Status Feedback and Closed-Loop Correction: Continuously obtain the latest pouring status through the multi-task visual perception module and feed it back to the pouring decision module to dynamically correct the material feeding and vibration parameters to prevent over-vibration, under-vibration or uneven material distribution. S9. Pouring Completion Judgment and Process End: When the pouring decision module determines that the pouring has reached the preset completion conditions, it stops the material feeding and vibration operation, outputs the pouring completion signal, and ends the current pouring process.
10. The precise material placement and intelligent pouring control method according to claim 9, characterized in that, In step S3, the initial pouring stage is in the state of no material pile, the intermediate filling stage is in the state of small material pile / large material pile, and the near-complete stage is in the state of not fully poured / fully poured. In step S8, the opening and closing angle of the material hopper gate and the frequency of the vibrator are dynamically closed-loop corrected by real-time collected information on the opening and closing angle of the material hopper gate and the liquid level, so as to realize the adaptive control of the pouring process.