Impact load and response intelligent prediction system and method based on machine vision

Abstract

The invention belongs to the field of explosion load and response prediction, and particularly relates to an impact load and response intelligent prediction system and method based on machine vision. An impact load and response intelligent prediction system based on machine vision comprises an image data preparation module, an image recognition and segmentation module, a numerical calculation module and a load and response prediction module, and the image data preparation module obtains, marks and enhances an image data set required by a target detection and tracking model; the image recognition and segmentation module recognizes and segments splashing fragments and dynamically tracks the splashing fragments; the numerical calculation module generates a sample database required for training the machine learning model; and the load and response prediction module outputs a forward prediction model, a reverse prediction model and a prediction result. According to the prediction system and method, computer vision, a numerical calculation method and machine learning are fused together, and the problems that a traditional method is low in efficiency and cannot conduct reverse prediction according to fragment distribution after impact explosion are solved.

Description

Technical Field

[0001] This invention belongs to the field of prediction of explosive load and response, specifically a machine vision-based intelligent prediction system and method for impact load and its response. Background Technology

[0002] Brittle materials such as concrete, rock, and glass exhibit significant fragmentation under explosive impact loads. The shock wave generated by the explosion propels fragments at high speeds, triggering secondary damage effects. The size, mass, and velocity distribution of the fragments are directly related to the explosion energy. Based on the correlation between explosive impact loads and fragment distribution characteristics, establishing a rapid prediction method for explosive impact loads and fragmentation response has extremely important engineering application value in civil engineering, mining engineering, military protection engineering, and materials science. Specifically, given the explosive yield and location, it is possible to positively predict and assess the degree of damage to structures and the range of fragmentation. In civil and mining engineering, it is possible to assess the safety of blasting demolition schemes or the danger range of flyrock in mine blasting. In military protection engineering, it is possible to design fortifications and armored structures that can effectively resist explosive impacts and fragmentation attacks. In public safety, it is possible to predict the hazard radius of hazardous facilities such as chemical plants and gas stations in explosion accidents. Similarly, by analyzing the distribution of debris and structural damage traces left at the scene, information such as the explosive yield and location can be deduced. In accident investigations, this can help determine the cause and responsibility of the explosion. In terms of weapon effectiveness assessment, the damage effect of weapons can be inverted through post-war on-site assessment. In terms of forensic engineering analysis, it can provide scientific basis for judicial and insurance appraisals.

[0003] Because explosive loads are highly transient, it is difficult to measure explosive fragments in real time through experimental means. Currently, the analysis is mostly carried out by manually collecting fragments after the event, such as by using sieving to count the percentage of fragments in each mass range.

[0004] Numerical methods offer advantages such as flexible design, controllable process, high repeatability, convenient data acquisition, and high cost-effectiveness, making them a valuable supplement to experimental methods. Currently, numerical methods for brittle material impact explosion problems mainly include mesh-dependent methods such as the Eulerian method, the Lagrangian method, and hybrid methods combining the advantages of both, as well as meshless methods represented by the smoothed particle flow method, the material point method, and the discrete element method. However, due to the large scale of real-world engineering problems, current computer hardware and algorithms require considerable time to complete the calculations, failing to meet the needs of real-time prediction applications in engineering practice.

[0005] Furthermore, traditional experimental and numerical methods are suitable for analyzing and solving forward problems, that is, predicting response results based on the properties of load, materials and structures, but cannot make inverse predictions based on the distribution of debris after an impact explosion. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent prediction system and method for impact load and its response based on machine vision. This prediction system integrates computer vision, numerical calculation methods and machine learning, and solves the problems of low efficiency and inability to make inverse predictions based on the distribution of debris after impact explosion in traditional methods.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: On one hand, this invention provides an intelligent prediction system for impact load and its response based on machine vision, including an image data preparation module, an image recognition and segmentation module, a numerical calculation module, and a load and response prediction module. The image data preparation module is used to acquire, label, and enhance the image dataset required by the target detection and tracking model, providing a data foundation for the entire system. Its input is source video, and its output is a training dataset. The image recognition and segmentation module is used to identify and segment splashed debris and dynamically track it to extract key features of the debris. Its input is the training dataset from the image data preparation module, and its output is a trained image detection and segmentation machine learning model. The numerical calculation module is used to generate a sample database required for training the machine learning model. Its input is determined parameter indicators and value space, and its output is a sample database of numerical calculation results. The load and response prediction module is used to train a fast surrogate model using the sample database generated by the numerical calculation module, realizing a forward prediction model that predicts debris distribution from impact load and a reverse prediction model that infers impact load from debris distribution. Its input is the sample database generated by the numerical calculation module and new query data, and its output is the forward prediction model, the reverse prediction model, and the prediction results.

[0008] Preferably, the image data preparation module includes a data acquisition module, a data extraction and segmentation module, and a data annotation module. The source videos used in the data acquisition module include real impact explosion experiment videos captured by a preset high-speed camera and numerical simulation videos simulated by discrete element method (DEM). The data extraction and segmentation module extracts individual frame images from the source videos and divides all frame images into training sets, validation sets, and test sets. The data annotation module annotates the images in the training sets, validation sets, and test sets and generates corresponding label files.

[0009] More preferably, the frame images provided in the data extraction and segmentation module include fragmented motion images of the target structure in video frames at different time steps; the data labeling module uses Labelme software to generate JSON format label files from the frame images provided in the data extraction and segmentation module. The label files contain the precise polygonal outline coordinates and category label information of the target structure in the image. Converting the JSON format label files into TXT format label files completes the dataset required by the image recognition and segmentation module.

[0010] Preferably, the image recognition and segmentation module includes an image detection and segmentation machine learning model and a high-precision target tracking algorithm. The image detection and segmentation machine learning model extracts and analyzes the static features of the fragments, including the number of fragments, particle size distribution, and shape factor. The high-precision target tracking algorithm performs real-time motion detection on the splashed fragments and extracts the dynamic features of the fragments, including the motion trajectory, velocity vector field, and acceleration of each fragment.

[0011] More preferably, the image detection and segmentation machine learning model adopts the YOLOv13 model. The model weights selected for training the YOLOv13 model network are from the yolov13n.pt file. The "Detect" mode in the YAML configuration file is replaced with the "Segment" mode. The maximum number of training epochs is set to 2000 epochs, including an early stopping mechanism. The image size is set to 640×640 pixels, and the batch size is 16 by default. Using the trained YOLOv13 model, accurate detection and segmentation of objects in images / videos can be achieved from a two-dimensional plane perspective. At the same time, the fragment contour mask of the segmented image in any frame is output for subsequent statistical calculation of equivalent particle size distribution features.

[0012] More preferably, the high-precision target tracking algorithm includes the BoT-SORT algorithm, ByteTrack algorithm, DeepSORT algorithm, SORT algorithm, and Hungarian algorithm; the high-precision target tracking algorithm, combined with the trained YOLOv13 model, realizes the tracking of the motion trajectory of the target object's explosion impact fragments in the video from a two-dimensional plane perspective, thereby calculating the instantaneous velocity of each fragment in the two-dimensional direction at any frame of the video, providing data support for subsequent velocity distribution feature calculation.

[0013] Preferably, the parameters determined in the numerical calculation module include explosion load parameters, material and structural parameters and their respective value ranges; the numerical calculation methods used include smooth particle flow method, material point method, and discrete element method; the numerical calculation results include fragment number, equivalent particle size distribution, fragment mass distribution, velocity in each axis, load impact energy, characteristic dimensions in each axis, and spatial coordinates.

[0014] Preferably, the forward prediction model in the load and response prediction module uses the explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables; the inverse prediction model in the load and response prediction module uses the overall distribution characteristics of the fragments as input variables and the explosion load parameters, material and structural parameters as output variables; the forward prediction model and the inverse prediction model adopt one of the following: multilayer perceptron, XGBoost model, and LightGBM model.

[0015] On the other hand, the present invention provides a machine vision-based intelligent impact load and response prediction method, comprising the following steps: S1. Visual perception model construction: The image data preparation module acquires the source video of the broken object, extracts individual frame images from the source video, and divides all frame images into training set, validation set and test set in an 8:1:1 ratio; Labelme software is used to label the images of training set, validation set and test set and generate corresponding JSON label files, and then the JSON label files are converted into TXT format files by Python program; Configure the training environment for the image detection and segmentation machine learning model and obtain the YOLOv13 model weight parameters. Use the YOLOv13 model, train it using the training set obtained in step S1, and use the validation set to monitor performance and adjust hyperparameters. Use the pre-trained weight yolov13n.pt. The key hyperparameter settings are as follows: maximum training epochs are 2000, and early stopping mechanism is enabled. The input image size is set to 640×640 pixels, and the batch size is 16. After training, evaluate it on the test set and output the optimal weight file best.pt, which is the trained YOLOv13 segmentation and detection model. By combining the trained YOLOv13 model with a high-precision target tracking algorithm, a final visual perception model that can achieve fragment detection, segmentation, and tracking is obtained. S2. Construction of Physical Prediction Model: The numerical calculation module determines different explosion load parameters, material and structural parameters and their respective value ranges, uses high-precision numerical calculation methods to obtain detailed physical process data for each simulation, extracts input-output data pairs from them, and combines all input-output data pairs into a sample database. A machine learning model is trained using a sample database to learn the complex nonlinear relationship from input to output, resulting in a well-trained prediction model. The forward prediction model uses explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables. The reverse prediction model uses the overall distribution characteristics of the fragments as input variables and explosion load parameters, material and structural parameters as output variables. S3. Predictive Applications: In the reverse prediction application: (1) Obtain the impact video to be detected; (2) Use the visual perception model obtained in S1 to detect, segment and track the impact video, and output the fragment contour mask and fragment displacement; (3) Calculate the equivalent particle size distribution characteristics of the fragment field based on the fragment contour mask, and calculate the velocity distribution characteristics of the target structure at the moment of breakage under impact load based on the fragment displacement; (4) Use the obtained equivalent particle size distribution characteristics and velocity distribution characteristics as input information, and output the explosion load parameters, material and structural parameters through the reverse prediction model trained in S2. In the application of positive prediction: (1) Using the impact load information and material and structural parameters as input information, the positive prediction model trained in S2 is called to directly output the overall distribution characteristics of the fragments.

[0016] The beneficial effects of this invention are as follows: 1. This invention achieves reverse prediction by tracing the cause from the effect. Traditional methods, if only debris remains at the scene, make it extremely difficult to deduce the impact load at that time, requiring extensive reliance on expert experience and numerous trial calculations, resulting in low accuracy and efficiency. The reverse prediction model established in this invention can directly and quickly deduce the most likely impact load parameters based on the debris distribution structure identified by the machine vision recognition module. This has immeasurable value in fields such as accident investigation, weapon effectiveness assessment, and structural safety inversion.

[0017] 2. A significant leap in efficiency has been achieved through the numerical calculation module and the load and response prediction module. Existing technologies using numerical simulation methods such as finite element analysis require substantial computational resources and time. In contrast, this application trains a fast proxy model by pre-calculating a sample library, solidifying complex physical processes into a near-instantaneous input-output mapping relationship. Whether it's forward prediction of load results or reverse calculation of load parameters, it can be completed in seconds or even milliseconds, enabling rapid decision-making, real-time monitoring, and online evaluation.

[0018] 3. The image data preparation module and image recognition and segmentation module improve the automation and objectivity of state perception and feature extraction. Utilizing a trained machine learning model, video recognition, measurement, and tracking can be completed automatically, continuously, and with high precision. Key features such as position, speed, size, shape, and motion aggregation are accurately extracted, transforming this unstructured visual information into structured data suitable for numerical analysis, thus avoiding errors and fatigue associated with manual recognition.

[0019] 4. This system provides a powerful tool for safety design, accident prevention, and emergency response planning. Impact load tests (such as impact and explosion) are typically costly, dangerous, and non-reproducible. This application, through numerical simulation and surrogate models, enables numerous "virtual experiments" to be conducted in the digital space, significantly reducing reliance on physical experiments and saving costs and time. It allows for the prediction and evaluation of structural responses under extreme loads without damaging the structure or requiring hazardous experiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall working architecture of the prediction system of this invention; Figure 2 This is a schematic diagram of a falling hammer test simulation model; Figure 3 These are images showing the crushing effect of samples at different time steps; Figure 4 This is a schematic diagram of the fragment tracking effect of the YOLOv13 model; Figure 5 This is a schematic diagram of the fragment mask for model segmentation; Figure 6 It is a curve comparing the accuracy of the actual simulation speed and the model detection speed; Figure 7 It is a graph showing the accuracy comparison between the model training curve and the actual impact load and the model predicted load; Figure 8 This is a flowchart of the prediction method in this invention. Detailed Implementation

[0021] 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.

[0022] like Figure 1As shown, this invention provides an intelligent prediction system for impact load and its response based on machine vision, including an image data preparation module, an image recognition and segmentation module, a numerical calculation module, and a load and response prediction module. The image data preparation module is used to acquire, label, and enhance the image dataset required by the target detection and tracking model, providing a data foundation for the entire system. Its input is source video, and its output is a training dataset. The image recognition and segmentation module is used to identify and segment splashed debris and dynamically track it to extract key features of the debris. Its input is the training dataset from the image data preparation module, and its output is a trained image detection and segmentation machine learning model. The numerical calculation module is used to generate a sample database required for training the machine learning model. Its input is determined parameter indicators and value space, and its output is a sample database of numerical calculation results. The load and response prediction module is used to train a fast surrogate model using the sample database generated by the numerical calculation module, realizing a forward prediction model that predicts debris distribution from impact load and a reverse prediction model that infers impact load from debris distribution. Its input is the sample database generated by the numerical calculation module and new query data, and its output is the forward prediction model, the reverse prediction model, and the prediction results.

[0023] The image data preparation module includes a data acquisition module, a data extraction and segmentation module, and a data annotation module. The source videos used in the data acquisition module include real impact explosion experiment videos captured by a preset high-speed camera and numerical simulation videos simulated by discrete element method (DEM). The data extraction and segmentation module extracts individual frame images from the source videos and divides all frame images into training sets, validation sets, and test sets. The data annotation module annotates the images in the training set, validation set, and test set and generates corresponding label files.

[0024] The frame images provided by the data extraction and segmentation module in the image data preparation module include fragmented motion images of the target structure at different time steps in the video frames. The data annotation module uses Labelme software to generate JSON format tag files from the frame images provided by the data extraction and segmentation module. The tag files contain the precise polygonal contour coordinates and category label information of the target structure in the image. The JSON format tag files are then converted into TXT format tag files to obtain the dataset required by the image recognition and segmentation module.

[0025] The image recognition and segmentation module includes an image detection and segmentation machine learning model and a high-precision target tracking algorithm. The image detection and segmentation machine learning model extracts and analyzes the static features of the fragments, including the number of fragments, particle size distribution, and shape factor. The high-precision target tracking algorithm performs real-time motion detection on the splashed fragments and extracts the dynamic features of the fragments, including the motion trajectory, velocity vector field, and acceleration of each fragment.

[0026] The image detection and segmentation machine learning model uses the YOLOv13 model. The model weights selected for training the YOLOv13 model network are from the yolov13n.pt file. The "Detect" mode in the YAML configuration file is replaced with the "Segment" mode. The maximum number of training epochs is set to 2000 epochs, including an early stopping mechanism. The image size is set to 640×640 pixels, and the batch size is 16 by default. Using the trained YOLOv13n-seg model, accurate detection and segmentation of objects in images / videos are achieved from a two-dimensional plane perspective. At the same time, the fragment contour mask of the segmented image in any frame is output for subsequent statistical calculation of equivalent particle size distribution features.

[0027] The high-precision target tracking algorithms include the BoT-SORT algorithm, ByteTrack algorithm, DeepSORT algorithm, SORT algorithm, and Hungarian algorithm. These high-precision target tracking algorithms, combined with a trained YOLOv13 model, enable the tracking of the motion trajectory of debris from an explosion in a video from a two-dimensional plane perspective. This allows for the calculation of the instantaneous velocity of each debris in the two-dimensional direction at any frame of the video, providing data support for the subsequent calculation of velocity distribution characteristics.

[0028] The parameters determined in the numerical calculation module include the explosion load parameters, material and structural parameters and their respective value ranges; the numerical calculation methods used include the smooth particle flow method, the material point method, and the discrete element method; the numerical calculation results include the number of fragments, the equivalent particle size distribution, the fragment mass distribution, the velocity in each axis, the load impact energy, the characteristic dimensions in each axis, and the spatial coordinates.

[0029] The forward prediction model in the load and response prediction module takes the explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables; the inverse prediction model in the load and response prediction module takes the overall distribution characteristics of the fragments as input variables and the explosion load parameters, material and structural parameters as output variables; the forward prediction model and the inverse prediction model adopt one of the following: multilayer perceptron, XGBoost model, and LightGBM model.

[0030] like Figure 8 As shown, this invention also provides a machine vision-based intelligent prediction method for impact load and response. Taking a brittle limestone sample as an example: under a drop hammer impact load, the sample breaks and generates splashed fragments. The fragments are tracked and segmented using the above method, and the impact load is predicted inversely based on the detected particle size and velocity distribution. The specific prediction process is as follows: The rock sample drop hammer impact test was simulated using the discrete element method (DEM) software Rocky DEM. A schematic diagram of the experiment is shown below. Figure 2As shown, the Tavares Breakage model was selected to describe the rock fragmentation behavior. The drop hammer height was fixed at 0.75m. By changing the drop hammer mass, 30kg, 60kg, 90kg, and 120kg were selected in this embodiment to adjust the impact energy. The rock sample was a cylindrical limestone, and its dimensions were achieved by adjusting the length-to-diameter ratio. The physical parameters of the limestone sample and the drop hammer model are shown in Table 1, and the contact parameters between the sample and the boundary are shown in Table 2. The fragmentation effects of the sample at different time steps are shown in Table 2. Figure 3 As shown.

[0031] Table 1 Physical parameters of limestone samples and falling weight Table 2 Rock Specimen - Boundary Contact Parameters S1. Visual perception model construction: Multiple sets of video and image data of rock fracturing process in drop hammer impact simulation test were exported from Rocky DEM, and all frame images were divided into training set, validation set and test set in an 8:1:1 ratio. The images in the training set and validation set were labeled with polygonal outer contours using Labelme software and corresponding JSON label files were generated. Then, the JSON label files were converted into TXT format files required by YOLOv13 model using Python program.

[0032] Simultaneously, a portion of complete, undisassembled simulation video files is retained to verify the reliability of the YOLO model after training. Data such as fragment size distribution, velocity distribution, load, and rock sample type are exported for subsequent result verification or use in reverse prediction models.

[0033] Configure the training environment for the image detection and segmentation machine learning model and obtain the YOLOv13 model weight parameters. Use the YOLOv13 model, train it using the training set, and use the validation set to monitor performance and adjust hyperparameters. Use the pre-trained weight file yolov13n.pt. The key hyperparameter settings are as follows: maximum training epochs are 2000, and early stopping is enabled. The input image size is set to 640×640 pixels, and the batch size is 16. After training, evaluate it on the test set and output the optimal weight file best.pt, which is the trained YOLOv13 segmentation and detection model.

[0034] The image detection and segmentation machine learning model used in this embodiment is the YOLOv13 model. Obviously, other YOLO models can also be used. While inheriting the advantages of the YOLO series of algorithms, the YOLOv13 model introduces the HyperGraph Adaptive Enhancement (HyperACE) mechanism and the Full-Link Aggregation Allocation (FullPAD) paradigm, which significantly reduces the number of parameters and computational complexity while achieving a breakthrough in detection accuracy. It is very suitable for the detection, segmentation and tracking of small objects.

[0035] To further expand the dataset, the images obtained by the image data preparation module can be enhanced by methods such as rotation and flipping to increase the number of images.

[0036] By combining the trained YOLOv13 model with a high-precision target tracking algorithm, a final visual perception model that can achieve fragment detection, segmentation, and tracking is obtained. The high-precision target tracking algorithms mentioned above include, but are not limited to, BoT-SORT algorithm, ByteTrack algorithm, DeepSORT algorithm, SORT algorithm, Hungarian algorithm, etc.

[0037] S2. Construction of Physical Prediction Model: The numerical calculation module determines different explosion load parameters, material and structural parameters and their respective value ranges, uses high-precision numerical calculation methods to obtain detailed physical process data for each simulation, extracts input-output data pairs from them, and combines all input-output data pairs into a sample database.

[0038] A machine learning model is trained using a sample database to learn the complex nonlinear relationship from input to output, resulting in a well-trained prediction model. The forward prediction model uses explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables. The reverse prediction model uses the overall distribution characteristics of the fragments as input variables and the explosion load parameters, material and structural parameters as output variables.

[0039] The parameters determined in the numerical calculation module include the explosion load parameters, material and structural parameters and their respective value ranges; the numerical calculation methods used include the smooth particle flow method, the material point method, and the discrete element method; the numerical calculation results include the number of fragments, the equivalent particle size distribution, the fragment mass distribution, the velocity in each axis, the load impact energy, the characteristic dimensions in each axis, and the spatial coordinates.

[0040] Forward and backward prediction models can employ multilayer perceptrons, XGBoost models, LightGBM models, etc. In this embodiment, an MLP prediction model, namely multilayer perceptron, is used.

[0041] S3. Predictive Applications: This embodiment takes reverse prediction application as an example: (1) Obtain the impact video to be detected; (2) Use the visual perception model obtained in S1 to detect, segment and track the impact video, and output the fragment contour mask and fragment displacement; (3) Calculate the equivalent particle size distribution characteristics of the fragment field according to the fragment contour mask, and calculate the velocity distribution characteristics of the target structure at the moment of breakage under impact load according to the fragment displacement; (4) Use the obtained equivalent particle size distribution characteristics and velocity distribution characteristics as input information, and output the explosion load parameters, material and structural parameters through the reverse prediction model trained in S2. In the application of positive prediction: (1) Using the impact load information and material and structural parameters as input information, the positive prediction model trained in S2 is called to directly output the overall distribution characteristics of the fragments.

[0042] In this embodiment, the visual perception model consists of the YOLOv13 model combined with the BoT-SORT multi-object tracking algorithm, which can detect, segment, and track rock fragments generated in a drop hammer simulation experiment. The detection and tracking results are as follows: Figure 4 As shown, the segmented motion fragment contour mask is as follows: Figure 5 As shown; the velocity distribution characteristics of the target structure at the moment of breakage under impact load are calculated using fragment displacement, and the velocity calculation accuracy comparison curve is shown in Figure 1. Figure 6 As shown; the training loss curve of the inverse prediction model and its prediction accuracy for impact loads are compared as follows. Figure 7 As shown.

[0043] The above embodiments detail the process of the inverse prediction model for predicting impact load parameters, while the core principle of the forward prediction model for predicting fragment distribution characteristics is the same as that in the above embodiments. Based on the inverse analysis model, the forward prediction model can be constructed by replacing its input and output parameters, so it will not be elaborated here.

[0044] Multiple experiments have shown that the system of this invention has high accuracy and reliability in real-time prediction of the distribution and load of brittle material fragments in impact explosion scenarios, and has broad application prospects and market promotion value.

[0045] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A machine vision-based intelligent prediction system for impact loads and their responses, characterized in that: It includes an image data preparation module, an image recognition and segmentation module, a numerical calculation module, and a load and response prediction module. The image data preparation module is used to acquire, label, and enhance the image dataset required by the target detection and tracking model, providing a data foundation for the entire system. Its input is source video and its output is training dataset. The image recognition and segmentation module is used to identify and segment the splashed fragments and dynamically track them, thereby extracting the static and dynamic features of the fragments. Its input is the training dataset in the image data preparation module, and its output is the trained image detection and segmentation machine learning model. The numerical calculation module is used to generate a sample database required for training machine learning models. Its input is a defined parameter index and value space, and its output is a sample database of numerical calculation results. The load and response prediction module is used to train a fast surrogate model using the sample database generated by the numerical calculation module, and obtain a forward prediction model that predicts fragment distribution from impact load and a reverse prediction model that infers impact load from fragment distribution. Its input is the sample database generated by the numerical calculation module and new query data, and its output is the forward prediction model, the reverse prediction model and the prediction results.

2. The intelligent prediction system for impact load and its response based on machine vision according to claim 1, characterized in that: The image data preparation module includes a data acquisition module, a data extraction and segmentation module, and a data annotation module. The source videos used in the data acquisition module include real impact explosion experiment videos captured by a preset high-speed camera and numerical simulation videos simulated by discrete element method (DEM). The data extraction and segmentation module extracts individual frame images from the source videos and divides all frame images into training sets, validation sets, and test sets. The data annotation module annotates the images in the training sets, validation sets, and test sets and generates corresponding label files.

3. The intelligent prediction system for impact load and its response based on machine vision according to claim 2, characterized in that: The frame images provided in the data extraction and segmentation module include fragmented motion images of the target structure at different time steps in the video frames. The data annotation module uses Labelme software to generate JSON format tag files from the frame images provided in the data extraction and segmentation module. The tag files contain the precise polygonal outline coordinates and category label information of the target structure in the image. The JSON format tag files are then converted into TXT format tag files to obtain the dataset required by the image recognition and segmentation module.

4. The intelligent prediction system for impact load and its response based on machine vision according to claim 1, characterized in that: The image recognition and segmentation module includes an image detection and segmentation machine learning model and a high-precision target tracking algorithm. The image detection and segmentation machine learning model extracts and analyzes the static features of the fragments, including the number of fragments, particle size distribution, and shape factor. The high-precision target tracking algorithm performs real-time motion detection on the splashed fragments and extracts the dynamic features of the fragments, including the motion trajectory, velocity vector field, and acceleration of each fragment.

5. The intelligent prediction system for impact load and its response based on machine vision according to claim 4, characterized in that: The image detection and segmentation machine learning model adopts the YOLOv13 model. The model weights selected for training the YOLOv13 model network are from the yolov13n.pt file. The "Detect" mode in the YAML configuration file is replaced with the "Segment" mode. The maximum number of training epochs is set to 2000 epochs, including an early stopping mechanism. The image size is 640×640 pixels, and the batch size is 16 by default. Using the trained YOLOv13 model, the target structure is accurately detected and segmented in a two-dimensional plane view. At the same time, the fragment contour mask of the segmented image in any frame is output for subsequent statistical calculation of equivalent particle size distribution features.

6. The intelligent prediction system for impact load and its response based on machine vision according to claim 4, characterized in that: The high-precision target tracking algorithm includes the BoT-SORT algorithm, ByteTrack algorithm, DeepSORT algorithm, SORT algorithm, and Hungarian algorithm. The high-precision target tracking algorithm, combined with the trained YOLOv13 model, realizes the tracking of the motion trajectory of the explosive impact fragments of the target object in the video from a two-dimensional plane perspective, thereby calculating the instantaneous velocity of each fragment in the two-dimensional direction at any frame of the video, providing data support for the subsequent calculation of velocity distribution characteristics.

7. The intelligent prediction system for impact load and its response based on machine vision according to claim 1, characterized in that: The parameters determined in the numerical calculation module include explosion load parameters, material and structural parameters and their respective value ranges; the numerical calculation methods used include smooth particle flow method, material point method and discrete element method; the numerical calculation results include fragment number, equivalent particle size distribution, fragment mass distribution, velocity in each axis, load impact energy, characteristic dimensions in each axis and spatial coordinates.

8. The intelligent prediction system for impact load and its response based on machine vision according to claim 1, characterized in that: The forward prediction model in the load and response prediction module takes the explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables; the inverse prediction model in the load and response prediction module takes the overall distribution characteristics of the fragments as input variables and the explosion load parameters, material and structural parameters as output variables; the forward prediction model and the inverse prediction model adopt one of the following: multilayer perceptron, XGBoost model, and LightGBM model.

9. A prediction method using a machine vision-based intelligent prediction system for impact loads and their responses, as described in any of the preceding claims, characterized in that: Includes the following steps: S1. Visual perception model construction: The image data preparation module acquires the source video of the broken object, extracts individual frame images from the source video, and divides all frame images into training set, validation set and test set in an 8:1:1 ratio; Labelme software is used to label the images of training set, validation set and test set and generate corresponding JSON label files, and then the JSON label files are converted into TXT format files by Python program; Configure the training environment for the image detection and segmentation machine learning model and obtain the YOLOv13 model weight parameters. Use the YOLOv13 model, train it using the training set, and use the validation set to monitor performance and adjust hyperparameters. Use the pre-trained weight file yolov13n.pt. The key hyperparameter settings are as follows: maximum training epochs are 2000, and early stopping mechanism is enabled. The input image size is set to 640×640 pixels, and the batch size is 16. After training, evaluate it on the test set and output the optimal weight file best.pt, which is the trained YOLOv13 segmentation and detection model. By combining the trained YOLOv13 model with a high-precision target tracking algorithm, a final visual perception model that can achieve fragment detection, segmentation, and tracking is obtained. S2. Construction of Physical Prediction Model: The numerical calculation module determines multiple sets of different explosion load parameters, material and structural parameters and their respective value ranges. It uses high-precision numerical calculation methods to obtain detailed physical process data for each set of simulations, extracts input-output data pairs from them, and combines all input-output data pairs into a sample database. A machine learning model is trained using a sample database to learn the complex nonlinear relationship from input to output, resulting in a well-trained prediction model. The forward prediction model uses explosion load parameters, material and structural parameters as input variables and the overall distribution characteristics of the fragments as output variables. The reverse prediction model uses the overall distribution characteristics of the fragments as input variables and explosion load parameters, material and structural parameters as output variables. S3. Predictive Applications: In the reverse prediction application: (1) Obtain the impact video to be detected; (2) Use the visual perception model obtained in S1 to detect, segment and track the impact video, and output the fragment contour mask and fragment displacement; (3) Calculate the equivalent particle size distribution characteristics of the fragment field based on the fragment contour mask, and calculate the velocity distribution characteristics of the target structure at the moment of breakage under impact load based on the fragment displacement; (4) Use the obtained equivalent particle size distribution characteristics and velocity distribution characteristics as input information, and output the explosion load parameters, material and structural parameters through the reverse prediction model trained in S2. In the application of positive prediction: (1) Using the impact load information and material and structural parameters as input information, the positive prediction model trained in S2 is called to directly output the overall distribution characteristics of the fragments.

Images