Construction of prediction model for laser irradiation of concrete block drop point and prediction method

Abstract

The application relates to the field of building structure maintenance and repair, and discloses a construction method and a prediction method of a laser irradiation concrete fragment landing point prediction model. The prediction model at least comprises a low-layer feature extraction network, a space self-adaptive feature extraction network based on internal convolution, and a multi-scale feature fusion network. The overall design follows a progressive feature extraction logic of 'from low-layer texture to high-layer semantics and from local details to global dependence'. Compared with a traditional convolution, the internal convolution generates a space transformation kernel through self-adaptation, thereby breaking the bottleneck of a traditional convolutional neural network in real-time performance and accuracy, significantly enhancing the modeling capability for the spatial specificity of laser energy distribution, reducing parameter redundancy, improving the model reasoning efficiency, and enabling fast reconstruction of fragment movement trajectories and landing point prediction, thereby providing technical support for active safety protection.

Description

Technical Field

[0001] This invention relates to the field of building structure maintenance and repair, and in particular to the construction and prediction method of a laser irradiation concrete fragment impact point prediction model. Background Technology

[0002] In the fields of laser processing and building demolition, high-energy laser irradiation technology for concrete structures is gradually becoming a cutting-edge application. However, this technology can cause concrete to burst during use, producing high-speed fragments. These fragments have initial velocities of 10-50 m / s, and their scattering direction exhibits strong randomness and spatial anisotropy, posing a serious threat to construction safety. Existing technologies lack the ability to predict the trajectory and landing point of fragments in real time; therefore, the protective measures adopted mostly rely on fixed safety distances and simple isolation, which are insufficient to meet the engineering requirements for efficient protection. Summary of the Invention

[0003] This invention provides a method for constructing and predicting the impact point of laser-irradiated concrete fragments, thus solving the aforementioned technical problems.

[0004] A first aspect of this invention provides a laser-irradiated concrete fragment impact point prediction model, comprising at least a low-level feature extraction network, a spatial adaptive feature extraction network, and a feature fusion network.

[0005] A low-level feature extraction network is used to extract low-level visual features from the input data and generate a low-level feature map. The input data includes time-stamped point cloud data of concrete surface, laser irradiation image data, and energy distribution image data.

[0006] The spatial adaptive feature extraction network is an improved RedNet-Involution network, which is used to dynamically generate the convolution kernel weights corresponding to each spatial location in the low-level feature map through multiple consecutively stacked involution modules, and to perform weighted aggregation on the low-level feature map to generate a spatial adaptive feature map. The spatial adaptive feature map is used to represent the differences in concrete surface features at each spatial location under different laser energy distributions during laser irradiation.

[0007] The feature fusion network is used to simultaneously extract spatial features of fragments at different distances from the spatial adaptive feature map through multiple involution branches of different scales. After dynamically adjusting the weights of each spatial feature through a learnable weight matrix, the features are weighted and fused to generate a fused multi-scale feature map, which is then used to predict the landing point of subsequent fragments.

[0008] A second aspect of this invention provides a method for constructing a model for predicting the impact point of laser-irradiated concrete fragments, comprising the following steps:

[0009] Step 1: Collect historical data of the target concrete structure and construct a training set. The historical data includes time-stamped point cloud data of the concrete surface, laser irradiation image data, and energy distribution image data.

[0010] Step 2: Preprocess the sample data of the training set and label the landing points of the fragments;

[0011] Step 3: Construct a prediction model for the impact point of laser-irradiated concrete fragments, and train the prediction model using the training set until the preset training target is achieved.

[0012] A third aspect of the present invention provides a method for predicting the impact point of laser-irradiated concrete fragments, which acquires current point cloud data of the concrete surface of the target concrete structure, current laser irradiation image data, and current energy distribution image data, and outputs the prediction result of the fragment impact point in real time based on the laser-irradiated concrete fragment impact point prediction model.

[0013] The beneficial effects of this invention are as follows: This invention provides a method for constructing and predicting the landing point of laser-irradiated concrete fragments. Compared with traditional convolution, involution overcomes the bottleneck of traditional convolutional neural networks in terms of real-time performance and accuracy by adaptively generating spatial transformation kernels. It significantly enhances the ability to model the spatial specificity of laser energy distribution, while reducing parameter redundancy and improving model inference efficiency. Thus, it can achieve rapid reconstruction of fragment movement trajectory and landing point prediction within a time limit of less than 100 milliseconds, providing technical support for active safety protection.

[0014] To make the above-mentioned objects, features and advantages of the invention more apparent and understandable, preferred embodiments of the invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the framework of a laser-irradiated concrete fragment impact point prediction model provided in one embodiment;

[0017] Figure 2 This is a schematic diagram of the feature fusion network architecture provided in one embodiment;

[0018] Figure 3 This is a flowchart illustrating a method for constructing a model to predict the impact point of laser-irradiated concrete fragments, as provided in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0020] It should be noted that, unless otherwise specified, the various features in the embodiments of this invention can be combined with each other, all of which are within the protection scope of this invention. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. Moreover, the terms "first," "second," and "third" used in this invention do not limit the data or execution order, but only distinguish identical or similar items with essentially the same function and effect.

[0021] To address the problems of computational complexity, limited two-dimensional information, and inability to model long-range spatial dependencies in traditional methods, one embodiment of this invention first proposes a laser-irradiated concrete fragment impact point prediction model based on an involution network. An improved RedNet-Involution network structure is used as the feature extraction backbone. The overall design follows a progressive feature extraction logic "from low-level texture to high-level semantics, from local details to global dependencies." The involution module consists of two parts: a feature dimensionality reduction module and a kernel weight generation module. The former reduces the number of feature channels to 1 / r of the original number through 1×1 convolutions, while the latter inputs the dimensionality reduction result into a kernel generator composed of lightweight activation functions, dynamically generating convolution kernel weights for different spatial locations according to the region, achieving spatially specific encoding. Detailed descriptions are provided below through specific embodiments.

[0022] like Figure 1 As shown, the laser irradiation concrete fragment impact point prediction model of Example 1 includes at least a low-level feature extraction network, a spatial adaptive feature extraction network, and a feature fusion network.

[0023] For example, a low-level feature extraction network is used to extract low-level visual features from the input data to generate a low-level feature map. The input data includes time-stamped synchronized point cloud data of the concrete surface, laser irradiation image data, and energy distribution image data. Specifically, in implementation, a laser scanning device is used to perform a high-density point cloud scan on the target concrete structure to obtain concrete surface point cloud data and construct a 3D surface model. Simultaneously, a high-speed industrial camera is used to record the laser irradiation process, obtaining laser irradiation image data, such as laser irradiation video data, to extract the movement trajectory of the fragments. The energy distribution image data can be in the depth channel (RGB-D or projected feature map), and the laser parameters and imaging device parameters are synchronized via time stamps to ensure the consistency of these input data.

[0024] First, the above data is input into a low-level feature extraction network, where preliminary feature encoding is performed using multi-layer convolution (Conv) and batch normalization (BatchNorm) modules. This stage is mainly responsible for extracting low-level visual features such as surface texture, edges, cracks, and aggregate distribution, providing a geometric basis for subsequent spatial adaptive analysis. The low-level visual features maintain high spatial resolution to achieve accurate position-level prediction in subsequent stages. The formula used is as follows:

[0025] ; ;

[0026] in, This is a low-level feature map. For the input image, Point cloud image of concrete surface, For laser irradiation images, This is an image of energy distribution.

[0027] like Figure 1 As shown, the spatial adaptive feature extraction network is an improved RedNet-Involution network, which is used to dynamically generate the convolution kernel weights corresponding to each spatial location in the low-level feature map through multiple consecutively stacked involution modules, and to perform weighted aggregation on the low-level feature map to generate a spatial adaptive feature map. The spatial adaptive feature map is used to represent the differences in concrete surface features at each spatial location under different laser energy distributions during laser irradiation.

[0028] Here, the spatial adaptive feature extraction network is the core of the entire prediction model. In this stage, multiple consecutively stacked involute modules are introduced. Unlike traditional convolution, the involute network can dynamically generate convolution kernel weights for each spatial location of the input feature map, capturing the differences in energy distribution between regions and the inhomogeneity of the internal structure of concrete.

[0029] As those skilled in the art know, the core idea of ​​involution is to dynamically generate position-dependent convolutional kernels for each spatial location, thereby achieving adaptive weight allocation at the spatial level. Let the input feature map be... The size of the involution kernel is Involution in position The output is defined by Formula 2:

[0030] ;

[0031] in Indicates Centered Neighborhood set For position For neighborhood location The generated involution weight vector (which can be assigned by group) has the following symbol: This represents the element-wise product by channel (or the element-wise product under grouping / channel mapping). If grouping is used... and dimensionality reduction The common implementation is as follows: first, for Perform down-channel mapping , by a lightweight kernel generating function Generate each position Weights, as shown in Formula 3:

[0032]

[0033] Therefore, position The output can be written in matrix form (by expanding the neighborhood into a vector), as shown in Formula 4:

[0034]

[0035] in It is the first The above formula reveals that the involution at each spatial location actually performs a position-dependent weighted sum, with the weights generated by the local feature network, thereby achieving spatial adaptation.

[0036] Compared to regular convolution ( Input channel Output channel ): Parameters are Involution typically employs dimensionality reduction ratio r and grouping. and kernel-generated small networks (parameters) The overall parameters are approximately given by Formula 5:

[0037]

[0038] Compared to traditional convolution, the computational complexity of involution is as shown in Formula 6:

[0039]

[0040] The computational complexity of standard convolution is given by Formula 7:

[0041]

[0042] In the input channel Dimensionality reduction Under these conditions, the number of involution parameters is reduced by approximately The number of times, and because the kernels of involution are generated positionally and typically through a smaller generative network, is significantly higher. (For example The globally learnable parameters are much smaller than those of an equivalent conventional convolutional layer. The multiplication and addition complexity during actual inference can be effectively controlled through grouping and dimensionality reduction, which is more conducive to real-time (<100ms) inference requirements.

[0043] Involution is learnable and end-to-end trainable, providing an example of loss. Regarding kernel generation parameters Gradient representation. Let position Output As shown in Equation 4, the loss is relative to the position. The gradient of the involution kernel is given by formula 8:

[0044] ,

[0045] Then, by the chain rule, the gradient of the parameters of the generating function g is given by Equation 9:

[0046] ,

[0047] The above equation shows that learning involution relies on the interaction between input neighborhood features and upper-layer error signals, which is similar to conventional convolution. However, the weights are generated by the local feature network, thus better capturing spatial specificity.

[0048] Involution mathematically separates kernel generation from neighborhood weighting, allowing the response at each spatial location to be adaptively determined by the local input. Its dimensionality reduction and grouping strategies significantly reduce the number of global parameters and operating costs, while the backpropagation formula ensures end-to-end training of the involution module. By combining spatial specificity and channel sharing mechanisms, this model can learn the local differences in the material response to laser energy in different regions, enabling the network to possess location-dependent feature description capabilities. The spatially adaptive feature map output at this stage not only preserves local structural information but also enhances the perception of crack evolution and fragmentation trends in the laser-affected area, laying a feature foundation for subsequent multi-scale fusion.

[0049] like Figure 1As shown, the feature fusion network is used to simultaneously extract spatial features of fragments at different distances in the spatial adaptive feature map through multiple involution branches of different scales, and then perform weighted fusion after dynamically adjusting the weights corresponding to each spatial feature through a learnable weight matrix to generate a fused multi-scale feature map, so as to predict the subsequent fragment landing point based on the multi-scale feature map.

[0050] Specifically, the feature fusion network inputs spatially adaptive feature maps in parallel into three involution branches of different scales, such as... Figure 2 As shown, the 1×1 involute kernel is used to capture local details of fragments (such as crack edges and the trajectory of small particles); the 3×3 involute kernel is used to extract medium-range contextual information (such as scattering direction and spatial correlation of fragment groups); and the 5×5 involute kernel is responsible for capturing global spatial dependencies (such as the relationship between laser energy diffusion and the overall concrete structure).

[0051] like Figure 2 As shown, the outputs of each branch are dynamically adjusted by the learnable weight matrices (W1, W2, W3) and then enter the weighted fusion layer for element-wise summation to obtain the fused multi-scale feature map, as shown in Equation 10:

[0052]

[0053] Each of these The calculations all follow the above formula for position-dependent generation and neighborhood weighting.

[0054] The fusion mechanism in this embodiment can fully aggregate multi-level information while maintaining consistent spatial resolution, achieving a balance between detailed and global features. This allows for subsequent fragment landing point prediction based on the multi-scale feature map, enhancing the model's robust prediction capability for fragment landing points in complex scenarios.

[0055] In a preferred embodiment, such as Figure 1 As shown, the laser-irradiated concrete fragment impact point prediction model further includes a regression branch and a confidence estimation branch. The fused high-dimensional feature map obtains the impact point prediction result through the regression branch and the confidence estimation branch. Specifically, the regression branch determines the existence of fragments based on the multi-scale feature map and outputs the predicted impact point of the fragments. The predicted result includes the three-dimensional coordinates or probability distribution of the fragment impact point. The regression branch is trained using a binary cross-entropy loss function to supervise the existence prediction of fragments and an Euclidean distance loss function to supervise the prediction of the three-dimensional coordinates or probability distribution of the fragment impact point. The confidence estimation branch is used to output the confidence level of the predicted result, thereby evaluating the uncertainty and reliability of the prediction result.

[0056] The laser-irradiated concrete fragment impact point prediction model in the above embodiments utilizes a deep learning framework. By constructing an involution neural network that combines spatial specificity and channel sharing mechanisms, it achieves high-precision modeling and prediction of fragment movement trajectories. The network model employed is a multi-scale involution structure, capable of simultaneously processing spatial information around the fragment at different scales such as 1×1, 3×3, and 5×5. Through parallel computation, it dynamically generates adaptive involution kernel weights corresponding to spatial positions, thereby capturing the non-uniformity of energy distribution and local feature changes on the concrete surface during laser irradiation. Furthermore, the generation process of this involution kernel can employ a bottleneck structure incorporating linear transformation and nonlinear activation, significantly reducing the number of parameters and computational complexity while maintaining feature extraction capabilities. This achieves high-precision prediction of the impact point of laser-irradiated concrete fragments, balancing real-time performance, robustness, and deployability, providing reliable intelligent algorithm support for on-site laser processing safety monitoring.

[0057] Based on the model structure design of the laser-irradiated concrete fragment impact point prediction model, another embodiment of the present invention also provides a method for constructing the laser-irradiated concrete fragment impact point prediction model, such as... Figure 3 As shown, it includes the following steps:

[0058] Step 1: Collect historical data of the target concrete structure and construct a training set. The historical data includes historical concrete surface point cloud data synchronized with timestamps, historical laser irradiation image data, and historical energy distribution image data.

[0059] Step 2: Preprocess the sample data of the training set and label the landing points of the fragments;

[0060] Step 3: Construct a prediction model for the impact point of laser-irradiated concrete fragments, and train the prediction model using the training set until the preset training target is achieved.

[0061] In one specific embodiment, step 2 preprocesses the sample data of the training set and labels the fragment landing points, specifically as follows:

[0062] Step 201: Voxel mesh downsampling is performed on the concrete surface point cloud data, and normal vector estimation is performed on the downsampled point cloud data to generate optimized point cloud data with normal vector information. For example, Open3D and PCL libraries can be used to downsample the point cloud of an average of 4.1 million points per frame to about 120,000 points, and the normal vector structure can be estimated by the KNN (k=30) method to reflect the geometric features of the concrete surface.

[0063] Step 202, in terms of video processing, extract the debris scattering trajectory from the laser irradiation image data, and perform three-dimensional coordinate projection in conjunction with camera calibration to generate and label the debris landing point coordinates. For example, the debris scattering trajectory is extracted from a 1920×1080, 240fps laser irradiation video using the OpenCV frame differencing method, and three-dimensional coordinate projection is achieved in conjunction with camera calibration, resulting in 276 landing point data points.

[0064] Step 203: Store the optimized point cloud data, the fragment scattering trajectory, and the fragment landing point coordinates in a unified world coordinate system, with the labels using the extended COCO format.

[0065] In a preferred embodiment, the preprocessing further includes a data augmentation step, specifically: using random rotation, illumination perturbation, Perlin noise, and / or local gradient enhancement methods to increase sample data corresponding to differences in concrete surface changes under different observation angles, surface reflections, surface roughness environments, and different laser irradiation temperature differences. For the laser-irradiated area, the Sobel operator is first used to calculate the horizontal and vertical gradients in the image, then edge information within the area is extracted, and the surface reflection or brightness changes caused by laser irradiation are analyzed. Then, the gradient coefficient is amplified by 2-3 times to enhance the texture and reflection changes in the area, allowing the model to focus more on these areas with significant changes during sample training. For cracks or temperature difference areas on the concrete surface, edge detection algorithms, such as Canny edge detection, are used to identify crack areas, or infrared images are used to detect temperature difference areas. Then, normal vector difference calculation is used to calculate the change in normal vectors between adjacent points in the point cloud data to capture surface details, making the surface changes caused by cracks and temperature differences more prominent, helping the model better identify subtle features related to the fragment impact point.

[0066] During the implementation phase, this gradient-enhanced data will be used to generate new training samples, thereby improving the model's ability to perceive laser-irradiated areas, cracks, and areas of temperature difference changes, enhancing its prediction accuracy for complex surface features, and thus improving the prediction model's ability to perceive surface changes caused by different lighting, angles, materials, and temperature differences after laser irradiation, i.e., improving the generalization performance of the prediction model.

[0067] In a preferred embodiment, the training process in step 3 employs a joint loss function, as shown in Equation 11:

[0068]

[0069] in Used for fragment detection and presence determination The regression error is the coordinate of the landing point. This ensures the smoothness and continuity of the predicted spatial distribution.

[0070] Compared to traditional monitoring methods based on physical simulation and two-dimensional image analysis, the prediction model trained through the above embodiments, through the unique spatial specificity mechanism of involution operations, can accurately capture the spatial variation characteristics of fragment scattering trajectories caused by the non-uniformity of laser energy distribution and the anisotropy of concrete materials, thus achieving high-precision modeling of fragment motion behavior. The dynamic weight generation capability of the involution kernel enables the network to adapt to energy diffusion patterns under different working conditions, overcoming the limitations of the spatial invariance assumption of traditional convolutional neural networks. Simultaneously, the channel-sharing design significantly reduces the number of model parameters, improves computational efficiency, and meets the timeliness requirements of real-time on-site monitoring. The prediction model constructed by the method of this invention, through a multi-scale feature fusion architecture, effectively integrates local details and global contextual information, improving the prediction accuracy of fragment landing points in complex environments compared to traditional methods.

[0071] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0072] This invention also provides a method for predicting the impact point of laser-irradiated concrete fragments. The method involves real-time acquisition of current concrete surface point cloud data, current laser irradiation image data, and current energy distribution image data of the target concrete structure, and real-time output of the predicted impact point of the fragments based on the laser-irradiated concrete fragment impact point prediction model constructed using the above method.

[0073] In addition, in another preferred embodiment, the prediction results of deep learning can be seamlessly integrated with the three-dimensional point cloud model of the target concrete structure. An intuitive heat map of the probability distribution of the landing point is generated based on the predicted three-dimensional coordinates of the fragment landing point, providing direct and visualized decision support for safety early warning. For example, it can be used as a trigger reference for automatic braking systems and personnel avoidance devices, significantly enhancing the safety protection capability of the laser processing process and having important engineering application value.

[0074] This invention is not limited to the descriptions and embodiments herein; therefore, other advantages and modifications can be readily realized by those skilled in the art. Thus, without departing from the spirit and scope of the general concept defined by the claims and their equivalents, this invention is not limited to the specific details, representative devices, and illustrative examples shown and described herein. Those skilled in the art, guided by the teachings of this invention, can make many modifications without departing from the spirit and scope of the claims, all of which are within the protection of this invention.

Claims

1. A method for constructing a model for predicting the impact point of laser-irradiated concrete fragments, characterized in that, Includes the following steps: Step 1: Collect historical data of the target concrete structure and construct a training set. The historical data includes time-stamped point cloud data of the concrete surface, laser irradiation image data, and energy distribution image data. Step 2: Preprocess the sample data of the training set and label the landing points of the fragments; Step 3: Construct a prediction model for the impact point of laser-irradiated concrete fragments, and train the prediction model using the training set until the preset training target is achieved. The laser-irradiated concrete fragment impact point prediction model includes at least a low-level feature extraction network, a spatial adaptive feature extraction network, and a feature fusion network. The low-level feature extraction network is used to extract low-level visual features from the input data and generate a low-level feature map. The input data includes time-stamped point cloud data of concrete surface, laser irradiation image data, and energy distribution image data. The spatial adaptive feature extraction network is an improved RedNet-Involution network, which is used to dynamically generate the convolution kernel weights corresponding to each spatial position in the low-level feature map through multiple consecutively stacked involution modules, and to perform weighted aggregation on the low-level feature map to generate a spatial adaptive feature map. The spatial adaptive feature map is used to represent the differences in concrete surface features at each spatial position under different laser energy distributions during laser irradiation. The feature fusion network is used to simultaneously extract spatial features of fragments at different distances from the spatial adaptive feature map through multiple involution branches of different scales. After dynamically adjusting the weights corresponding to each spatial feature through a learnable weight matrix, the feature fusion is performed to generate a fused multi-scale feature map, which is then used to predict the landing point of subsequent fragments. The feature fusion network includes at least three involution branches of different scales. The first involution branch is used to extract local detail features of fragments, the second involution branch is used to extract context information within a preset range, and the third involution branch is used to extract global spatial dependencies. The local detailed features of the fragments include crack edges and / or the movement trajectory of small particles; the preset range context information includes the direction of fragment scattering and / or the spatial correlation of fragment groups; and the global spatial dependency includes the dependency between laser energy diffusion and the overall concrete structure.

2. The construction method according to claim 1, characterized in that, The laser-irradiated concrete fragment impact point prediction model also includes a regression branch and a confidence estimation branch. The regression branch is used to determine whether a fragment exists based on the multi-scale feature map and output the prediction result of the fragment landing point. The prediction result includes the three-dimensional coordinates or probability distribution of the fragment landing point. The confidence estimation branch is used to output the confidence level of the prediction result.

3. The construction method according to claim 1 or 2, characterized in that, The low-level visual features include the texture, edges, cracks, and / or aggregate distribution characteristics of the concrete structure surface.

4. The construction method according to claim 1, characterized in that, The sample data of the training set is preprocessed, and the landing points of the fragments are labeled, specifically as follows: The point cloud data of the concrete surface is downsampled using voxel mesh, and the normal vector of the downsampled point cloud data is estimated to generate optimized point cloud data with normal vector information. The debris scattering trajectory is extracted from the laser irradiation image data, and a three-dimensional coordinate projection is performed in conjunction with camera calibration to generate and mark the debris landing point coordinates; The optimized point cloud data, the fragment scattering trajectory, and the fragment landing point coordinates are all stored in the world coordinate system.

5. The construction method according to claim 4, characterized in that, The preprocessing of the sample data in the training set also includes a data augmentation step, specifically: using random rotation, illumination perturbation, Perlin noise and / or local gradient enhancement methods to increase the sample data corresponding to the differences in concrete surface changes under different observation angles, surface reflection, surface roughness environments and different laser irradiation temperature differences.

6. A method for predicting the impact point of laser-irradiated concrete fragments, characterized in that, The system acquires the current point cloud data of the concrete surface, the current laser irradiation image data, and the current energy distribution image data of the target concrete structure, and outputs the prediction results of the fragment landing point in real time based on the laser irradiation concrete fragment landing point prediction model constructed according to any one of claims 1-5.

Images