Slip form construction template residue identification method and automatic cleaning system

Abstract

The invention relates to the technical field of slip-form construction, and provides a slip-form construction template residue identification method and an automatic cleaning system.The slip-form construction template residue identification method comprises the steps that image information of a template surface unit area and infrared thickness information of residues are collected; respectively preprocessing the image information and the infrared thickness information; the method comprises the steps of recognizing image information of the surface of a template based on a pre-constructed image recognition model, outputting a residue type, a predicted area and a predicted thickness, obtaining a measured thickness based on infrared thickness information, carrying out weighted fusion on the measured thickness and the predicted thickness to obtain a calibrated thickness, and outputting the calibrated thickness. And calculating the cleaning grade of the unit region based on the residue prediction area and the calibration thickness, and judging the cleaning mode of the template in combination with the residue type and the cleaning grade, thereby realizing the automatic cleaning of the template residues.

Description

Technical Field

[0001] This invention relates to the field of slipform construction technology, and in particular to a method for identifying slipform template residue and an automatic cleaning system. Background Technology

[0002] Slipform construction is a construction process that uses a hydraulic system to lift the formwork to achieve continuous concrete pouring. It is necessary to clean the concrete residue adhering to the formwork surface regularly. If the residue is not removed, it will harden and cause misalignment of the formwork joints, leading to quality problems such as grout leakage and honeycomb pitting on the structural surface. Inverted formwork structures are prone to demolding difficulties due to incomplete cleaning.

[0003] Existing cleaning methods lack automation, relying primarily on manual scraping with tools like scrapers and wire brushes, or simply high-pressure water washing. Limited by worker experience and physical strength, the cleaning quality is inconsistent. Furthermore, large formwork areas are difficult to cover quickly manually, and manual cleaning accounts for 15%-20% of the construction cycle. The formwork structure requires frequent repositioning, posing high safety risks for high-altitude operations. Simply using high-pressure water or mechanical scraping cannot flexibly adjust to residues with varying adhesion strengths, such as loose cement slurry and solidified aggregates, easily leading to over-cleaning or incomplete cleaning. For complex structures like formwork, precise adaptation is difficult. Summary of the Invention

[0004] To address the problem of low intelligence in template cleaning in existing technologies, the first objective of this invention is to provide a method for identifying template residue in slipform construction, comprising:

[0005] Image information of the template surface unit area and infrared thickness information of the residue are collected;

[0006] Image information and infrared thickness information are preprocessed separately;

[0007] Image information of the template surface is identified based on a pre-built image recognition model, and the residue type, predicted area and predicted thickness are output. The measured thickness is obtained based on infrared thickness information. The measured thickness and the predicted thickness are weighted and fused to obtain the calibration thickness. The cleaning level of the unit area is calculated based on the residue predicted area and the calibration thickness. The cleaning method of the template is determined by combining the residue type and the cleaning level.

[0008] Specifically, the image recognition model is based on the MobileNetV3-Small backbone network and includes:

[0009] Input layer;

[0010] The feature extraction layer uses the depthwise separable convolution module of MobileNetV3, and a batch normalization layer and a Hard-Swish activation function are added after the 3rd, 5th and 7th convolution layers.

[0011] The feature fusion layer concatenates the feature maps from layers 4 and 9 through channels;

[0012] The task output layer includes a classification branch and a regression branch. The classification branch compresses the channel through a 1×1 convolution and outputs the residue type probability using the Softmax activation function. The regression branch uses a linear activation function and outputs the residue thickness prediction value through a 1×1 convolution.

[0013] Specifically, a channel and spatial attention module is set after the feature fusion layer. The channel attention weight focuses on the residual feature channel, and the spatial attention weight suppresses the template background interference area.

[0014] Specifically, the classification branch is equipped with an anchor frame adaptive adjustment module and an intersection-union ratio matching mechanism, and the anchor frame size is adjusted based on the size of the residue.

[0015] Specifically, the training method for the image recognition model is as follows:

[0016] Construct a dataset and divide it into a training set, a validation set, and a test set;

[0017] Add L2 regularization and Dropout layers, and set pre-trained weights;

[0018] The first 5 layers of the backbone network are frozen, and the feature fusion layer and output layer are trained with pre-trained weights. After multiple rounds of training, all layers are unfrozen and fine-tuned with a smaller learning rate.

[0019] The training is conducted in stages: the first stage trains only the classification branch, the second stage trains only the regression branch, and the third stage trains both branches together.

[0020] Specifically, the objective function L of the image recognition model is:

[0021] L = α × Lcls + β × Lreg;

[0022] In the formula, Lcls is the classification loss, which adopts weighted cross-entropy loss and sets cross weights according to the difference in the number of residue samples; α is the classification loss weight; Lreg is the regression loss, which adopts weighted root mean square error and sets weights according to the residue thickness level; and β is the regression loss weight.

[0023] Specifically, data perturbation and morphological enhancement are added to the dataset. The data perturbation includes ±20% brightness disturbance, ±15% contrast adjustment, and Gaussian noise with a variance of 0-0.01. The morphological enhancement includes rotation of 0-90°, scaling of 0.8-1.2x, and shearing within the range of 10%.

[0024] Specifically, the cleaning levels include light adhesion, moderate adhesion, and heavy adhesion.

[0025] The residue thickness of the light adhesion grade is ≤0.5mm and its proportion in the unit area is <10%;

[0026] The thickness of the residue with the medium adhesion grade is between 0.5 mm and 2 mm, or the proportion in the unit area is between 10% and 30%.

[0027] The residue of the heavy adhesion grade is thicker than 2 mm and accounts for more than 30% of the unit area. A second objective of this invention is to provide an automatic cleaning system for slipform construction templates, using the aforementioned slipform construction template residue identification method, including:

[0028] The moving unit has a guide rail at the top of the template, and the moving unit moves circumferentially in the guide rail;

[0029] An image acquisition module is disposed in the moving unit, which scans the template surface circumferentially and acquires image information of the template surface unit area;

[0030] An infrared measurement module is installed in the moving unit to detect the infrared thickness information of the residue in the unit area on the template surface;

[0031] The edge computing module acquires image information and infrared thickness information of the unit area, preprocesses the image information and infrared thickness information respectively, identifies the type of residue on the template surface, predicted area and predicted thickness based on the pre-built image recognition model, and classifies the cleaning level into light adhesion, medium adhesion and heavy adhesion based on the predicted area of ​​residue and the calibrated thickness obtained by weighted fusion of predicted thickness and measured thickness.

[0032] The cleaning unit is suspended at the bottom of the mobile unit and moves circumferentially on the template surface to clean the surface residue of the unit area based on the cleaning level and residue type.

[0033] Specifically, the cleaning unit includes a negative pressure suction nozzle, a brush roller, and a spray head installed in the mobile unit, and a pressure sensor is provided in the brush roller.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention uses an image recognition model to identify the type, area, and thickness of residue in the template, and obtains a calibrated thickness by weighted fusion of the residue thickness measured by infrared. Based on the predicted area and thickness of the residue, the cleaning level is divided. The cleaning unit can adjust different cleaning methods according to the cleaning level, realizing automated cleaning of template residue. The cleaning time per session is shortened from 20-30 minutes to 3-5 minutes, which not only ensures the residue removal rate but also reduces the template damage rate, thereby extending the service life of the template.

[0036] 2. In the image recognition model, this invention adds batch normalization (BN) layers and Hard-Swish activation functions after the convolution of layers 3, 5, and 7 in the feature extraction layer to avoid gradient vanishing and enhance feature expression capabilities. In order to address the problem of low recognition accuracy, a "shallow-deep feature fusion" mechanism is introduced in the feature fusion layer. The feature maps of layer 4 (shallow layer, which retains small-sized residue details) and layer 9 (deep layer, which extracts semantic information of residues) are fused by channel splicing. The task output layer adopts a dual-branch structure to output the residue type and the predicted value.

[0037] 3. The cleaning unit uses a negative pressure suction nozzle, brush roller and spray head. For light adhesion, the negative pressure adsorption mode is used to clean the residue. For moderate adhesion, mechanical brushing combined with low pressure water mist is used to clean the residue. For heavy adhesion, strong brushing combined with water jet is used to clean the residue. It is suitable for cleaning needs under different adhesion conditions. Attached Figure Description

[0038] Figure 1 This is a flowchart of the template residue identification process of the present invention;

[0039] Figure 2 This is a schematic diagram of the automatic cleaning system of the present invention.

[0040] Reference numerals: 1. Guide rail; 2. Brush roller; 3. Nozzle; 4. Lifting sling; 5. Moving unit; 6. CCD camera; 7. Edge computing module; 8. Template. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The application principles of this invention will be further described below with reference to the accompanying drawings.

[0042] As shown in the figure, this invention provides a method for identifying concrete residue on slipform construction templates and an automatic cleaning system. By using an image recognition model and an infrared measurement module to collaboratively identify the area and thickness of concrete residue on the template surface, and combining an integrated multi-mode composite cleaning module with high-pressure water jet, flexible steel wire brush roller, and negative pressure adsorption, it solves the problems of low efficiency, high template damage, and incomplete residue removal in traditional manual cleaning. It is especially effective for complex structures such as large templates and inverted formwork, achieving efficient template cleaning, ensuring the continuity of slipform construction, and extending the service life of the template.

[0043] The method for identifying template residue in slipform construction is as follows: image information of unit areas on the template surface and infrared thickness information of the residue are collected; the image information and infrared thickness information are preprocessed respectively; the image information of the template surface is identified based on a pre-built image recognition model, and the residue type, predicted area and predicted thickness are output; the cleaning level of the unit area is calculated based on the predicted area and calibrated thickness of the residue; and the cleaning method of the template is determined by combining the residue type and the cleaning level.

[0044] A guide rail 1 is provided at the top of the template, and a moving unit 5 is arranged in the guide rail. The moving unit 5 moves circumferentially within the guide rail 1. The moving unit 5 houses an image acquisition module for acquiring image information of the surface unit area of ​​the template 8 and an infrared thickness sensor for acquiring infrared thickness information of the residue. The image acquisition module can be a linear CCD camera 6, which can move horizontally along the guide rail and comprehensively scan the template surface circumferentially to acquire image information in real time. The CCD camera has a resolution of 0.1 mm / pixel. Alternatively, a LiDAR can be used for the image acquisition module to further improve the accuracy of 3D modeling. The infrared thickness sensor and image acquisition module employ an industrial-grade infrared displacement sensor based on the triangulation principle, calculating the residue thickness information by detecting the difference in infrared light reflection, with an accuracy of ±0.1 mm.

[0045] Edge computing module 7 collects image information and infrared thickness information, which are then recognized and weighted by an image recognition model to classify the cleaning level. The image recognition model adopts a lightweight convolutional neural network (CNN) architecture, with MobileNetV3-Small as the backbone network, adapted to the low computing power requirements of the edge computing unit, and combined with multi-task learning design. The specific structure includes:

[0046] The input layer receives image data acquired by a linear CCD camera, preprocesses the images, and uniformly adjusts the resolution to 224×224×3 to adapt to the MobileNet input specifications. The pixel values ​​are normalized to [0,1] to eliminate the influence of lighting differences on the model.

[0047] The feature extraction layer uses the depthwise separable convolution module of MobileNetV3 to reduce the number of parameters by more than 70% compared with traditional CNN, thereby improving inference speed. At the same time, batch normalization (BN) layers and hard-swish activation functions are added after the 3rd, 5th and 7th convolutional layers to avoid gradient vanishing and enhance feature representation ability.

[0048] The feature fusion layer introduces a "shallow-deep feature fusion" mechanism, fusing the feature maps of layer 4 (shallow layer, preserving details of small-sized residues) and layer 9 (deep layer, extracting semantic information of residues) through channel concatenation to solve the problem of low recognition accuracy for fine cement paste skins (size <2mm). After the feature fusion layer, CBAM (channel and spatial attention module) is added. Channel attention weights focus on residue feature channels (such as high grayscale difference channels of aggregate particles), while spatial attention weights suppress background interference areas such as template textures and scratches, increasing the model's attention weight to residue areas by more than 30%.

[0049] The task output layer has a dual-branch structure, including a classification branch and a regression branch. The classification branch compresses the number of channels to two using a 1×1 convolution, with the two channels representing two different residue types—cement paste skin and aggregate particles. It uses a Softmax activation function to output the residue type probability at a resolution of 7×7, corresponding to the original image's 224×224 grid division, achieving region-level type recognition. The regression branch outputs the predicted residue thickness (unit: mm) within the 7×7 grid using a 1×1 convolution, employing a linear activation function that directly corresponds to the thickness detection dimension of the infrared sensor. In the classification branch, an "anchor frame adaptive adjustment" mechanism is added. For small residues <5mm, a small anchor frame of 16×16 pixels is set; for aggregate particles >10mm, a large anchor frame of 64×64 pixels is set. This improves the positioning accuracy of residues of different sizes through an IOU (Intersection over Union) matching mechanism.

[0050] The image recognition model uses actual images of 200 slipform construction templates (taken by a linear CCD camera, resolution 0.1mm / pixel) and thickness data from infrared sensors recorded simultaneously as a dataset, acquiring a total of 15,000 valid images. The LabelMe tool is used to label "residue type (cement paste skin / aggregate particles)," "residue location (bounding box)," and "residue thickness (labeled based on infrared data, accuracy ±0.1mm)," generating PASCAL VOC format label files.

[0051] To address the challenges in identifying grain silo template residue due to its color being similar to the template, susceptibility to reflective interference, and significant size differences, this study simulates construction scene interference in the dataset and enhances the residue morphology by adding brightness perturbation (±20%), contrast adjustment (±15%), and Gaussian noise (variance 0-0.01) to cover different lighting environments such as cloudy days and strong light. Through random rotation (0-90°), scaling (0.8-1.2 times), and shearing (within 10%), the study simulates the irregular distribution of residue on the template surface (such as edge accumulation and local caking).

[0052] The dataset was divided into a training set (10,500 images), a validation set (3,000 images), and a test set (1,500 images) in a 7:2:1 ratio. The number of images at different thickness levels was balanced using the SMOTE algorithm (to avoid excessive lightly damaged images that could lead to model bias).

[0053] The training hardware uses an NVIDIA RTX 3090 GPU to accelerate feature computation, while the edge deployment uses an NVIDIA Jetson Nano with a computing power of 21 TOPS to meet real-time inference requirements.

[0054] The framework is based on TensorFlow 2.10 to build models, and uses TensorFlow Lite to quantize the model, converting 32-bit floating-point numbers into 8-bit integers, compressing the model size to 4.2MB, and improving inference speed by 2 times.

[0055] The training strategy for the model is as follows:

[0056] Transfer learning is used to first load the pre-trained weights of MobileNetV3 on the ImageNet dataset, freeze the first 5 layers of the backbone network, and train only the feature fusion layer and the output layer, resulting in fast convergence. After 10 epochs of training, all layers are unfrozen, and a smaller learning rate (initial learning rate 10) is applied. -4 →10 after thawing -5 Fine-tuning;

[0057] The training is divided into stages. In the first stage (10-25 rounds), only the classification branch is trained to optimize the identification of residue types. In the second stage (26-40 rounds), only the regression branch is trained to optimize the thickness prediction. In the third stage (41-60 rounds), both branches are trained together to balance the accuracy of the two types of tasks.

[0058] Regularization, adding L2 regularization (weight decay factor 10) -5 The addition of a Dropout layer (dropout rate=0.2) suppressed overfitting, and the validation set accuracy remained stable at over 95%.

[0059] Because the model needs to simultaneously perform the dual tasks of "residue type classification" and "thickness regression", a weighted combined objective function L is designed:

[0060] L=α×L cls +β×L reg ;

[0061] L clsFor classification loss (Classification_Loss): a weighted cross-entropy loss is adopted. For the difference in the number of samples of the two types of residues (cement paste samples account for 65%, aggregate particles account for 35%), weight coefficients are set (cement paste samples weight 0.35, aggregate particles weight 0.65) to avoid the model biased towards the majority class samples. α is the classification loss weight.

[0062] L reg The regression loss is calculated using a weighted root mean square error (RMSE). Weights are set according to the residue thickness level, such as 1.0 for light residue, 1.5 for moderate residue, and 2.0 for heavy residue. The focus is on optimizing the prediction accuracy of heavy residue (thickness > 2 mm) that has a greater impact on the cleaning decision. β is the regression loss weight.

[0063] Weighting coefficients α and β: determined by grid search of the validation set. In the preferred embodiment of the present invention, α=0.4 and β=0.6 to balance the loss contribution of the two types of tasks. If α is too large, it will lead to thickness prediction deviation; if β is too large, it will lead to misjudgment of residue type, such as misjudging aggregate particles as cement paste skin.

[0064] The edge computing unit uses an NVIDIA Jetson Nano edge controller (equipped with a Linux system, supporting real-time data processing). The computing process consists of 6 steps, with a total time of ≤0.5 seconds (meeting the real-time requirements of slipform construction).

[0065] S1: Receive data. Receive image data streams from the linear CCD camera (30fps, 2.5MB per frame) and analog signals from the infrared thickness sensor (100Hz) in real time via Ethernet. Data latency is controlled within 10ms.

[0066] S2: Preprocess the image information and infrared thickness information respectively. Resize (224×224), normalize (pixel value → [0,1]), and denoise (Gaussian filter, kernel size 3×3) the CCD image; perform moving average filtering (window size 5) on the infrared thickness signal to remove noise caused by construction vibration (filtered accuracy ±0.05mm).

[0067] S3: Model inference, using the deployed TensorFlow Lite quantization model to infer the preprocessed image, outputting the residue type distribution (type probability of each cell in a 7×7 grid) and the predicted residue thickness (thickness of each cell in a 7×7 grid), with an inference time ≤ 0.3 seconds / frame;

[0068] S4: Result fusion. The "weighted fusion" strategy is adopted to combine the predicted thickness obtained by image inference and the measured thickness obtained by infrared thickness sensor to calculate the calibration thickness. Calibration thickness = γ × predicted thickness + θ × calibration thickness, where γ and θ are the weights of predicted thickness and calibration thickness, respectively. γ is preferably 0.6 and θ is preferably 0.4. The error after fusion is ≤0.08mm, which effectively improves the accuracy of thickness determination.

[0069] S5: Determine the cleaning level by calculating the percentage of residual area pixels across the entire template. Match the cleaning level based on the "thickness and area percentage": For light adhesion, the residual thickness is ≤0.5mm and its percentage in the unit area is <10%; for moderate adhesion, the residual thickness is between 0.5mm and 2mm, or its percentage in the unit area is between 10% and 30%; for heavy adhesion, the residual thickness is >2mm and its percentage in the unit area is >30%.

[0070] S6: Output command, which sends the cleaning level and residue type signals to the cleaning unit via RS485 bus, controls it to switch to the corresponding mode to clean the unit area, and uploads the cleaning level signal to the construction monitoring platform.

[0071] The cleaning unit cleans residue according to the cleaning level and residue type signals. The cleaning unit includes a negative pressure suction nozzle, brush roller 2, and spray head 3 installed in the moving unit. It is suspended at the bottom of the moving unit by a lifting sling 4. The lifting sling 4 controls the vertical movement of the cleaning unit and moves it circumferentially along the guide rail surface with the moving unit, allowing it to move to a designated position. A flip guide rail is provided in the guide rail 1, enabling the cleaning unit to switch between the inside and outside of the template 8, cleaning residue from both the inner and outer surfaces of the template 8. The negative pressure suction nozzle is 50mm wide and has a suction power of ≥5kPa, used to adsorb loose residue on the surface; the brush roller 2 is a flexible steel wire brush roller with a surface covered with nylon and steel wire composite bristles, a diameter of 100mm, and a rotation speed of 0-200rpm. The brush roller 2 is equipped with a pressure sensor, which can adjust the pressure between the brush roller and the template surface according to different cleaning levels. In addition, for special templates, such as stainless steel mirror templates, the brush roller can use a polyurethane soft brush instead of a steel wire brush to avoid scratching the surface; the nozzle 3 is connected to the water pump through a pipe and can spray water mist of 3-8Mpa. The water mist pressure sprayed from the nozzle can be adjusted according to the cleaning level.

[0072] Cement slurry has a loose texture, low hardness, is easy to peel off, and has weak adhesion. Aggregate particles are hard, contain solidified concrete particles, and have strong adhesion. The tool parameters and auxiliary methods to be used at this level are determined according to the type of residue to ensure thorough cleaning without damaging the formwork.

[0073] For slight adhesion, if it is cement slurry skin, negative pressure adsorption is used to clean the residue. The strip negative pressure nozzles set below the unit area use high-speed airflow to peel off the loose residue on the template surface and transport it to the dust collection box through the pipeline to avoid dust. The width of the negative pressure nozzle is 50mm and the suction power is ≥5kP.

[0074] If it is aggregate particles, on the basis of negative pressure adsorption, start the brush roller at a low speed (50-100rpm) for slight friction, and use 1-2MPa low pressure water mist to soften the particles in the gaps to avoid residue.

[0075] For moderate adhesion, use mechanical brushing combined with low-pressure water mist to clean the residue. If it is cement slurry skin, rotate the brush roller at a low speed (100-150rpm) to break the adhesion of the residue. Spray the nozzle evenly with medium-pressure water mist at 3MPa in the contact area of ​​the brush roller to moisten the loose slurry skin and prevent the brush bristles from generating sparks. Control the brush roller pressure at 20-30N.

[0076] If it is aggregate particles, the brush roller rotates at a high speed (150-200 rpm) to break the adhesion of residue. The nozzle sprays a medium-high pressure water mist of 4-5 MPa evenly in the contact area of ​​the brush roller. The water mist impacts and breaks up the coagulation and adhesion between particles, avoiding the brush roller only grinding the surface and leaving the core.

[0077] For heavy adhesion, use a strong brush combined with water jet to clean the residue. If it is cement slurry skin, although it is thick, it is loose in texture. Use a high-speed (200rpm) brush roller with 5-6MPa high-pressure water mist. The water mist disperses the large pieces of slurry skin, and the brush roller helps to remove the residue.

[0078] If it is aggregate granules, because they are completely solidified and have extremely strong adhesion, a 5-8MPa high-pressure water mist first breaks up the solidified granule structure. The spray angle is 60°, and the brush roller rotates at a high speed (200rpm). Equipped with a pressure sensor, the contact force between the brush roller and the template surface is adjusted. The contact force is no more than 50N. The brush roller simultaneously removes blocky residues, avoiding scratches on the template surface caused by friction from hard particles. It is especially suitable for the treatment of stubborn residues in molded structures.

[0079] It employs a powerful brushing technique combined with water jet cleaning to remove residue. The brush roller rotates at high speed and is equipped with a pressure sensor to automatically adjust the contact force between the brush roller and the template surface, ensuring that the contact force does not exceed 50N. The nozzle sprays water mist at a pressure of 5-8MPa and a spray angle of 60°. The high-pressure water impact breaks up solidified residue, and the brush roller removes blocky residues, avoiding hard impacts that could damage the template. It is especially suitable for handling stubborn residues in molded structures.

[0080] The bottom of the template is equipped with a guide channel. Wastewater from the cleaning process is collected through this channel and flows into a sedimentation tank containing flocculant. After being filtered by a centrifugal filter (50μm precision), it enters a pH adjustment tank (pH=7-8) and finally flows into a clear water tank, achieving water recycling. The guide channel and sedimentation tank are equipped with electric heating coils to maintain a water temperature above 5℃ in winter, preventing equipment freezing and ensuring normal cleaning of the large template in low-temperature environments.

[0081] After cleaning, spray water-based release agent evenly onto the template surface using atomizing nozzles, with the thickness controlled at 50±5μm, to reduce the subsequent concrete adhesion.

[0082] This automatic cleaning system for slipform construction significantly improves cleaning efficiency and construction quality while reducing safety risks. The system utilizes a guide rail system for automated operation, effectively avoiding the risks associated with manual high-altitude work. The cleaning time per cycle is reduced from 20-30 minutes to 3-5 minutes, shortening the overall construction cycle by 10%-15%, making it particularly suitable for the rapid cleaning of large formwork. Using an image recognition model combined with infrared thickness measurement, residue on the formwork surface is graded, ensuring a residue removal rate of ≥99.5% and controlling the flatness error of the concrete forming surface to within 2mm. This also reduces the formwork damage rate from 5% to below 0.1%, extending the formwork's service life by more than 30%. Due to thorough cleaning, the demolding quality of the cast-in-place structure is significantly improved.

[0083] The above-described 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 scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying formwork residue in slipform construction, characterized in that, include: Image information of the template surface unit area and infrared thickness information of the residue are collected; Image information and infrared thickness information are preprocessed separately; Image information of the template surface is identified based on a pre-built image recognition model, and the residue type, predicted area and predicted thickness are output. The measured thickness is obtained based on infrared thickness information. The measured thickness and the predicted thickness are weighted and fused to obtain the calibration thickness. The cleaning level of the unit area is calculated based on the residue predicted area and the calibration thickness. The cleaning method of the template is determined by combining the residue type and the cleaning level.

2. The method for identifying template residue in slipform construction according to claim 1, characterized in that, The image recognition model is based on the MobileNetV3-Small backbone network and includes: Input layer; The feature extraction layer uses the depthwise separable convolution module of MobileNetV3, and batch normalization layers and Hard-Swish activation functions are added after the 3rd, 5th and 7th convolution layers. The feature fusion layer concatenates the feature maps from layers 4 and 9 through channels; The task output layer includes a classification branch and a regression branch. The classification branch compresses the channel through a 1×1 convolution and outputs the residue type probability using the Softmax activation function. The regression branch uses a linear activation function and outputs the residue thickness prediction value through a 1×1 convolution.

3. The method for identifying template residue in slipform construction according to claim 2, characterized in that, The feature fusion layer is followed by a channel and spatial attention module. The channel attention weight focuses on the residual feature channel, and the spatial attention weight suppresses the template background interference area.

4. The method for identifying template residue in slipform construction according to claim 2, characterized in that, The classification branch is equipped with an anchor frame adaptive adjustment module and an intersection-union ratio matching mechanism, and the anchor frame size is adjusted based on the size of the residue.

5. The method for identifying template residue in slipform construction according to claim 2, characterized in that, The training method for the image recognition model is as follows: Construct a dataset and divide it into a training set, a validation set, and a test set; Add L2 regularization and Dropout layers, and set pre-trained weights; The first 5 layers of the backbone network are frozen, and the feature fusion layer and output layer are trained with pre-trained weights. After multiple rounds of training, all layers are unfrozen and fine-tuned with a smaller learning rate. The training is conducted in stages: the first stage trains only the classification branch, the second stage trains only the regression branch, and the third stage trains both branches together.

6. The method for identifying template residue in slipform construction according to claim 2, characterized in that, The objective function L of the image recognition model is: L=α×L cls +β×L reg ; In the formula, L cls For classification loss, a weighted cross-entropy loss is used, with cross-weights set based on the difference in the number of residual samples. α is the classification loss weight, and L... reg For regression loss, weighted root mean square error is used, with weights set according to the residue thickness level, and β is the regression loss weight.

7. The method for identifying template residue in slipform construction according to claim 5, characterized in that, The dataset includes data perturbation and morphological enhancement. The data perturbation includes ±20% brightness disturbance, ±15% contrast adjustment, and Gaussian noise with a variance of 0-0.

01. The morphological enhancement includes rotation of 0-90°, scaling of 0.8-1.2x, and shearing within the range of 10%.

8. The method for identifying formwork residue in slipform construction according to claim 1, characterized in that, The cleaning levels include light adhesion, moderate adhesion, and heavy adhesion. The residue thickness of the light adhesion grade is ≤0.5mm and its proportion in the unit area is <10%; The thickness of the residue with the medium adhesion grade is between 0.5 mm and 2 mm, or the proportion in the unit area is between 10% and 30%. The residue thickness of the heavy adhesion grade is >2mm, and its proportion in the unit area is >30%.

9. An automatic cleaning system for slipform construction templates, characterized in that, The method for identifying slipform construction template residue according to any one of claims 1-8 includes: The moving unit has a guide rail at the top of the template, and the moving unit moves circumferentially in the guide rail; An image acquisition module is installed in the moving unit and scans the template surface circumferentially to acquire image information of the template surface unit area; An infrared measurement module is installed in the moving unit to detect the infrared thickness information of the residue in the unit area on the template surface; The edge computing module acquires image information and infrared thickness information of the unit area, preprocesses the image information and infrared thickness information respectively, identifies the type of residue on the template surface, predicted area and predicted thickness based on the pre-built image recognition model, and classifies the cleaning level into light adhesion, medium adhesion and heavy adhesion based on the predicted area of ​​residue and the calibrated thickness obtained by weighted fusion of predicted thickness and measured thickness. The cleaning unit is suspended at the bottom of the mobile unit and moves circumferentially on the template surface to clean the surface residue of the unit area based on the cleaning level and residue type.

10. The automatic cleaning system for slipform construction templates according to claim 9, characterized in that, The cleaning unit includes a negative pressure suction nozzle, a brush roller, and a spray head installed in the mobile unit, and a pressure sensor is provided in the brush roller.

Images