A real-time structural analysis method and system based on neural network

Abstract

The present invention provides a neural network-based real-time structural analysis method and system, which are applied to the field of data processing technology. This application preprocesses and extracts structural features from manufacturing scene data to generate feature analysis data; performs real-time analysis of the vehicle structure based on a neural network model to generate real-time analysis feature parameters; performs outlier filtering on changes in structural features of adjacent frames based on a dynamic threshold algorithm to generate structural state change data; processes the structural state change data to generate structural health dynamic assessment data and structural anomaly features; processes the data based on preset structural safety determination rules and structural anomaly features to generate structural safety determination data; and processes the safety determination data and real-time analysis feature parameters to generate structural fault warning information and production adjustment control instructions.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a real-time structural analysis method and system based on a neural network. Background Art

[0002] In high-end automotive manufacturing, structural safety directly impacts vehicle performance and reliability, necessitating real-time monitoring of stress distribution, deformation trends, and material fatigue in key components (such as chassis and frame). Traditional methods rely on offline testing or single-point static analysis, which struggles to meet the real-time and comprehensive requirements under dynamic operating conditions.

[0003] Insufficient data processing capabilities: Traditional methods rely on manually designed rules to extract features from manufacturing scene data (such as welding images and sensor timing data), and are unable to cope with multi-dimensional data under complex working conditions (such as stress, deformation, and material attenuation).

[0004] Limited real-time analysis accuracy: Anomaly detection based on fixed thresholds or simple statistical models cannot adapt to dynamic working conditions (such as load changes and ambient temperature fluctuations), which can easily lead to misjudgments or missed detections.

[0005] Lack of closed-loop management: There is a lack of a full-process automation mechanism from data collection to production adjustment, and the linkage efficiency between fault warning and production control is low.

[0006] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0007] The purpose of this application is to provide a real-time structural analysis method and system based on a neural network, which at least to some extent overcomes the problems existing in the prior art. By acquiring data, preprocessing and feature extraction, with the help of neural networks and dynamic threshold algorithm analysis, evaluation data and abnormal features are generated, and scores and warning thresholds are generated in combination with safety rules. Finally, fault warnings and production adjustment instructions are generated, closed-loop management is achieved, and the real-time and accuracy of analysis are improved.

[0008] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0009] According to one aspect of the present application, a neural network-based real-time structural analysis method is provided, comprising: acquiring manufacturing scene data and basic information about the automobile structure; preprocessing and feature extraction of structural features in the manufacturing scene data to generate feature analysis data, including an edge contour feature point set, a feature vector probability distribution, abnormal feature filtering results, weight feature parameters, and structural key point feature values; performing real-time analysis of the automobile structure based on a neural network model to generate real-time analysis feature parameters, including a stress distribution prediction value, a deformation trend vector, and a material fatigue index; performing outlier filtering on changes in structural features of adjacent frames based on a dynamic threshold algorithm to generate structural state change data, including deformation displacement, stress fluctuation range, and material property decay rate; processing the structural state change data to generate structural health dynamic assessment data and structural abnormality features, including fatigue crack risk probability and key component failure warning indicators; processing based on preset structural safety judgment rules and structural abnormality features to generate structural safety judgment data, including a safety level score and a risk level warning threshold; processing the safety judgment data and real-time analysis feature parameters to generate structural fault warning information and production adjustment control instructions.

[0010] Another aspect of the present application is a neural network-based real-time structural analysis device, characterized in that it includes: an acquisition module for acquiring manufacturing scene data and basic automobile structure information; a processing module for preprocessing and extracting features from structural features in the manufacturing scene data to generate feature analysis data, including edge contour feature point sets, feature vector probability distributions, abnormal feature filtering results, weight feature parameters, and structural key point feature values; performing real-time analysis of the automobile structure based on a neural network model to generate real-time analysis feature parameters, including stress distribution prediction values, deformation trend vectors, and material fatigue indicators; performing abnormal value filtering on changes in structural features of adjacent frames based on a dynamic threshold algorithm to generate structural state change data, including deformation displacement, stress fluctuation range, and material performance decay rate; processing the structural state change data to generate structural health dynamic assessment data and structural abnormality features, including fatigue crack risk probability and key component failure warning indicators; processing based on preset structural safety judgment rules and structural abnormality features to generate structural safety judgment data, including safety level scores and risk level warning thresholds; processing the safety judgment data and real-time analysis feature parameters to generate structural fault warning information and production adjustment control instructions.

[0011] According to another aspect of the present application, an electronic device is characterized in that it includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-mentioned neural network-based real-time structural analysis method by executing the executable instructions.

[0012] The present application provides a neural network-based real-time structural analysis method and system, which acquires data, performs preprocessing and feature extraction, uses neural networks and dynamic threshold algorithm analysis to generate evaluation data and abnormal features, combines safety rules to generate scores and warning thresholds, and finally generates fault warnings and production adjustment instructions, realizing closed-loop management and improving the real-time and accuracy of analysis.

[0013] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of a real-time structural analysis method based on a neural network provided by an embodiment of the present application is shown;

[0015] Figure 2 A schematic structural diagram of a real-time structural analysis device based on a neural network provided in one embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0017] The following combination Figure 1 The following describes a real-time structural analysis method based on a neural network according to an exemplary embodiment of the present application. It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application are applicable to any applicable scenario.

[0018] In one embodiment, the present application also proposes a real-time structural analysis method and system based on a neural network. Figure 1 The following schematically shows a flow chart of a method for real-time structural analysis based on a neural network according to an embodiment of the present application. Figure 1 As shown, the method is applied to the server and includes:

[0019] S101, obtaining manufacturing scene data and basic vehicle structure information.

[0020] In one implementation, production line images are captured using industrial cameras in real time, capturing images of each step of the automotive production line. For example, during the welding process, images of the weld status of each vehicle body component are captured to obtain visual information such as the appearance characteristics and weld quality of the welded areas. Equipment operation videos are also recorded using cameras, such as a stamping press operating during the stamping of automotive parts, to analyze the press's operating status and the compliance of the stamping action.

[0021] Collect process parameter monitoring data, collect process parameter monitoring data from the sensors and control systems of production equipment, for example, in the painting process, collect process parameter data such as the spraying pressure, spraying speed, and paint temperature of the painting robot.

[0022] Basic vehicle structural information includes information on the fundamental properties, models, and analysis and prediction of each vehicle's structural components. It provides a comprehensive description of the vehicle's structure and provides a foundation for subsequent analysis. Obtaining a 3D component model involves using 3D modeling software or scanning equipment to obtain precise 3D models of each vehicle component. For example, a 3D model of a vehicle frame includes information such as its geometry, dimensions, and the connections between its components.

[0023] Collect material property parameters for the materials used in automotive structural components from material suppliers or relevant technical documentation, such as the yield strength, tensile strength, and elastic modulus of body steel panels. Obtain structural stress pre-analysis results. Use finite element analysis to perform a stress pre-analysis of the automotive structure and obtain structural stress pre-analysis results. For example, perform a pre-analysis of the stress distribution of the vehicle chassis under different load conditions to obtain stress pre-analysis results for various parts of the chassis.

[0024] S102 , preprocessing and feature extraction are performed on the structural features in the manufacturing scene data to generate feature analysis data.

[0025] In one embodiment, a set of edge contour feature points is extracted based on structural features in manufacturing scene data. Computer vision algorithms are used to process manufacturing scene images. Computer vision aims to enable computers to simulate the human visual system to acquire, process, and understand information from images or videos. Identifying structural edge contours is a key step. Edges are regions in an image with dramatic grayscale changes, representing the boundary information of an object's structure. By detecting edges, the shape and contour range of an object can be determined. First, the image is Gaussian filtered to remove noise, which can cause erroneous edge detection results. Next, the image's gradient magnitude and direction are calculated. The gradient magnitude reflects the intensity of grayscale changes in the image, while the gradient direction indicates the direction of grayscale changes. Non-maximum suppression is then performed to refine edges in the gradient magnitude image, retaining only points with the highest local gradient as edge points. Finally, a dual-threshold algorithm is used to determine true edges. Two thresholds are set: those above the high threshold are considered edges, those below the low threshold are considered non-edges, and those between the high and low thresholds are considered non-edges based on their connectivity with the edge at the high threshold.

[0026] During the automotive door weld quality inspection, an industrial camera captures images of the door weld area (resolution 1920×1080). Using the Canny algorithm, the image is first filtered with a 5×5 Gaussian kernel, then the Sobel gradient is calculated. Edges are extracted using a high threshold of 0.3 and a low threshold of 0.1. This yields a set of approximately 2000 points representing the edge contours of the door weld, characterizing the weld's shape and boundary.

[0027] The edge contour feature point set is calculated to generate a probability distribution of feature vectors. The edge contour point set is converted into feature vectors, and the probability of occurrence of each feature value is counted to form a probability distribution. The point set coordinates are converted into normalized feature vectors (such as gradient direction and curvature). Histogram statistics are used to calculate the frequency of occurrence of feature vectors in each interval. For the extracted weld edge point set, the curvature feature vector is calculated for each point. Curvature intervals are divided into intervals of 0.1. The number of occurrences of feature vectors in each interval is counted to generate a probability distribution histogram of the curvature feature vectors. For the aforementioned car door weld edge point set, the curvature value of each point is calculated (ranging from -2.5 to 3.0), and 60 intervals are divided into intervals of 0.1. The curvature value in the range of 0.5 to 0.6 appears 320 times, accounting for 16%. A probability distribution histogram of the curvature feature is generated (curvature interval on the horizontal axis, frequency on the vertical axis).

[0028] Perform outlier filtering on the eigenvector probability distribution to generate anomaly feature filtering results. After obtaining the eigenvector probability distribution, the data may contain some values ​​that deviate from the normal range. These outliers may be caused by measurement errors, equipment failures, special operating conditions, and other reasons. They can interfere with subsequent data analysis and model accuracy, so they need to be filtered to obtain anomaly feature filtering result that better reflects the true characteristics.

[0029] In statistics, if data follows a Gaussian distribution (normal distribution), then approximately 99.73% of the data will fall within the interval of the mean μ ± 3 times the standard deviation σ. In other words, the probability of data occurring outside this interval is extremely small. Therefore, a data point with an absolute value greater than μ + 3σ or less than μ - 3σ is considered an outlier. This method is simple and intuitive and suitable for cases where data approximately follows a Gaussian distribution. In industrial production, many measurement data approximately follow a Gaussian distribution under stable production conditions. For example, in automotive parts manufacturing, some dimensional measurement data and physical performance parameters can be detected for outliers using the 3σ principle. For example, in the weld curvature feature vector in the above example, if it is assumed to follow a Gaussian distribution, this principle can be used to identify anomalous curvature values.

[0030] When applying the 3σ principle to weld curvature feature vectors, the first step is to calculate the mean μ and standard deviation σ through statistical analysis of a large amount of normal weld curvature data. This may require collecting multiple sets of data during the production process to ensure data representativeness. Then, each newly collected weld curvature feature vector is evaluated individually. If the absolute value of a curvature value is greater than μ + 3σ, it is identified as an outlier. These outliers may indicate equipment vibration or abnormal welding parameters during the welding process. Filtering these outliers makes subsequent weld curvature-based quality assessment and process analysis more accurate and reliable. Using the isolation forest model, all weld curvature feature vector data is used as input to construct multiple isolation trees. The anomaly score for each feature vector is calculated, and outliers are identified and filtered based on a pre-set threshold. For the door weld curvature, the mean μ = 0.8 and the standard deviation σ = 0.4 are calculated. Based on the 3σ principle, curvature values ​​with an absolute value greater than 0.8 + 3 × 0.4 = 2.0 are filtered out. 150 abnormal points (such as points with curvature = 2.3) were filtered out from the original 2000 points, and the abnormal feature filtering results (1850 valid curvature points) were obtained.

[0031] A weighted calculation is performed on the filtered feature data to generate weighted feature parameters. After filtering out abnormal feature values, the remaining feature data has varying degrees of importance for describing and analyzing the object's structural characteristics. Weighted calculations can highlight the role of important features and downplay the influence of less important features, resulting in weighted feature parameters that more accurately reflect the object's essential characteristics and provide a more valuable basis for subsequent analysis and decision-making. Assigning weights to data based on feature importance quantitatively measures the value of each feature within the overall feature set. For example, in automotive manufacturing, for feature data reflecting welding quality, some features may have a greater impact on weld strength, while others may have a greater impact on weld appearance. Weighting can distinguish these varying degrees of influence and ultimately calculate a comprehensive weighted feature parameter.

[0032] Detailed explanation of the weighted model, Entropy Weight Method. Entropy is a concept used in information theory to measure the amount of information. In the Entropy Weight Method, the smaller the entropy value of a feature, the greater the amount of information it contains, and the higher the weight should be given in the decision-making process. The specific calculation process is as follows:

[0033] First, the filtered feature data is standardized to ensure comparability between different features. Calculate the entropy value of each feature , the formula is , where n is the number of samples, is the weight of the jth sample under the i-th feature. Finally, the weight is calculated based on the entropy value For example, in the filtered curvature feature vectors in this example, curvature features with lower entropy values ​​are more critical in describing weld structural characteristics and are therefore assigned higher weights. The entropy weighting method is suitable for situations where the data itself can reflect certain information differences and is widely used in comprehensive multi-metric evaluations. For example, in the comprehensive evaluation of automotive parts quality, multiple features such as dimensional accuracy, surface roughness, and hardness are weighted to obtain a comprehensive quality assessment indicator.

[0034] The attention mechanism draws on the attention principle of the human visual system, allowing the model to focus on important parts when processing data. In feature weighting scenarios, it automatically assigns weights by calculating the degree of correlation between features or their relevance to the target task. For example, in a deep learning-based automobile structural feature analysis model, the model uses internal computing units (such as matrix multiplication and activation functions) to learn the importance of each feature based on the input feature data, assigning different weights to different features. This mechanism is commonly used in the field of deep learning, especially when processing complex data such as images and text. For example, in image recognition tasks in automotive manufacturing, for models identifying surface defects in automotive parts, the attention mechanism can enable the model to focus more on defect-related features, such as irregular edges and color anomalies, while assigning lower weights to irrelevant features.

[0035] Taking the curvature feature vector of automobile welds as an example, there are multiple curvature feature vectors measured at different positions or under different process conditions. When calculating the weight using the entropy weight method, the entropy value of each curvature feature is calculated first. If a certain curvature feature changes less in different samples, its entropy value is small, which means that the information it provides is relatively more stable and more important, and it will get a higher weight in the weighted calculation. Then multiply each curvature feature vector by the corresponding weight, and perform accumulation and other operations to obtain the weighted curvature feature parameter. This parameter comprehensively considers the importance of different curvature features. Compared with simple feature data stacking, it can more effectively reflect the structural characteristics of the weld, and provide a more accurate basis for subsequent judgment of whether the weld quality is qualified and whether the welding process needs to be adjusted. The 1850 curvature points are re-statistically distributed according to the interval, and the entropy value of each interval is calculated. For example, the entropy value of the curvature interval of 0.5~0.6 =0.72, weight =(1-0.72) / [(1-0.72)+(1-0.85)+…]=0.22. Finally, the weighted curvature characteristic parameters are obtained (e.g., weighted average curvature = 0.78).

[0036] Based on the correlation between the feature points and the structural centroid, the feature parameters are weighted and adjusted to generate the feature values ​​of the key points of the structure. The feature analysis data consists of the edge contour feature point set, the probability distribution of the feature vector, the abnormal feature filtering results, the weighted feature parameters, and the feature values ​​of the key points of the structure. The weights of the feature parameters are adjusted according to the distance from the feature point to the structural centroid to highlight the characteristics of the key area. The correlation model uses the distance decay model, and the weight is inversely proportional to the distance. The formula is: in, is the distance from the feature point to the centroid, The distance between the weld edge feature point and the mass center of the welded part is calculated, and the feature points closer to the mass center (such as stress concentration areas) are given higher weights to generate the key point eigenvalues ​​of the structure, such as the weighted key curvature eigenvalues. The coordinates of the mass center of the door welded part are calculated (960,540), the coordinates of a certain curvature point are (900,500), and the distance is calculated. Pixel. The original weight of the point =0.22, adjusted weight =0.22×[10 / (72.11+10)]=0.027, generating the eigenvalues ​​of key points of the structure (such as key curvature = 0.65, corresponding to the area near the center of mass).

[0037] S103, performing real-time analysis on the vehicle structure based on the neural network model to generate real-time analysis feature parameters.

[0038] In one implementation, a stress distribution prediction model is used to process historical stress data from the vehicle structure and current sensor data to generate a predicted stress distribution value. Historical stress data is collected for a specific vehicle chassis structure, selecting 100 key stress monitoring points (S1-S100) covering load-sensitive locations such as the chassis longitudinal beams, cross beams, and suspension connection points. Under full load conditions (rated load of 4000 kg), the on-board stress sensor system continuously monitors and records stress data at each point at a sampling frequency of 100 Hz, collecting a total of 500 sets of complete operating condition data (each set containing stress values ​​for 100 points × 1000 time steps). Under full load conditions, the stress value of sensor S50 in the middle of the chassis longitudinal beam at a specific moment was 150 MPa, while the stress value of sensor S8 at the front end of the cross beam was 80 MPa.

[0039] Current sensor data acquisition, under a half-load condition (2000kg load), uses 20 core points (S1-S20, covering stress concentration areas) within the same sensor system to collect data in real time. The sampling frequency is synchronized to 100Hz, and a set of real-time stress values ​​for all 20 points is obtained at the current moment. Under the half-load condition, the real-time stress value for sensor S50 (the middle of the longitudinal beam) is 110MPa, and for sensor S8 (the front end of the crossbeam) it is 60MPa. Data normalization uses Min-Max Normalization, using the following formula: Where x is the original stress value, and are the minimum and maximum values ​​in the historical data set (e.g., the historical stress range is -50 MPa to 200 MPa). For S50 of 150 MPa under full load, the normalized value is (150 - (-50)) / (200 - (-50)) = 0.8; for S50 of 110 MPa under half load, the normalized value is (110 - (-50)) / 250 = 0.64.

[0040] The model uses a three-layer fully connected multilayer perceptron (MLP) with the following structure: input layer: 120 dimensions (100 historical points + 20 current points); hidden layer 1: 128 neurons, ReLU activation function; hidden layer 2: 64 neurons, ReLU activation function; hidden layer 3: 32 neurons, ReLU activation function; output layer: 10,000 neurons (corresponding to the stress value of a 100×100 grid).

[0041] The normalized 100-dimensional historical stress vector and the 20-dimensional current stress vector are concatenated into a 120-dimensional input vector, for example: ; Forward propagation: After the input vector is extracted through three hidden layers, a 10,000-dimensional vector is generated in the output layer and reshaped into a 100×100 stress distribution matrix.

[0042] The prediction results are shown below, visualized in a matrix: a 100×100 grid corresponds to the coordinates of the chassis's two-dimensional plane, and the predicted value at each grid point represents the stress intensity at that location. For example, the predicted stress value at the center of the chassis longitudinal beam (grid coordinate (50,50)) is 120 MPa (normalized to 0.76); the predicted value at the crossbeam-to-longitudinal beam junction (30,70) is 180 MPa. Outlier identification: The model also outputs a stress gradient change matrix. If the stress gradient at a point exceeds a threshold (e.g., the difference between adjacent points is >30 MPa), it is marked as a potential stress concentration area.

[0043] The model processes deformation data of the vehicle structure under different operating conditions, analyzes deformation trends using a time series neural network, and generates a deformation trend vector. An LSTM (Long Short-Term Memory) model is used, consisting of two LSTM layers (64 units each) and one fully connected layer. The input sequence length is 50 and the output dimension is 10. Deformation data for different operating conditions is collected under acceleration, braking, and cornering. The three-dimensional coordinates of key chassis points are acquired using a laser displacement sensor with a sampling frequency of 10 Hz, recording 1000 time steps for each operating condition. The model input sequence is the Z-axis displacement data (unit: mm) of the midpoint of the chassis crossbar under braking over 50 time steps, such as [0.12, 0.15, 0.18, ..., 0.25]. The model outputs a 10-dimensional deformation trend vector [0.05, 0.12, -0.03, ..., 0.08], where positive and negative numbers indicate the deformation direction and the magnitude indicates the trend strength.

[0044] The system processes material fatigue test data and online monitored material performance parameters to generate a material fatigue index based on a neural network fatigue damage accumulation model. A CNN-BLSTM hybrid model is used. The CNN layer (32 3×3 convolution kernels) extracts spatial features, while the BLSTM layer (2 layers, 64 cells each) processes temporal features. The output is a fatigue index between 0 and 1. Fatigue test data and SN curve data for aluminum alloy materials at different stress amplitudes, including 100 sets of cycle number-stress amplitude samples, are used. Online parameter monitoring collects real-time parameters such as material temperature, load spectrum, and vibration frequency.

[0045] The model input is historical fatigue test data for an aluminum alloy door component (e.g., 5 × 10^5 cycles at a stress amplitude of 200 MPa), combined with the current online monitoring temperature of 25°C and the dynamic load spectrum (amplitude of 180 MPa, frequency of 10 Hz). The model output is a material fatigue index of 0.35 (0 indicates no fatigue, 1 indicates full fatigue), indicating a current fatigue level of 35%.

[0046] The predicted stress distribution, deformation trend vector, and material fatigue index are processed to generate real-time analysis characteristic parameters. The predicted stress distribution (matrix), deformation trend vector (10-dimensional), and fatigue index (scalar) are normalized and then weighted fused using the weights (stress 0.5, deformation 0.3, fatigue 0.2). The normalized maximum stress point of the predicted stress distribution is 0.85. The normalized deformation trend vector is [0.1, 0.3, -0.1, ..., 0.2], with an L2 norm of 0.42. The fatigue index is 0.35. After fusion calculation, the real-time analysis characteristic parameters are 0.5 × 0.85 + 0.3 × 0.42 + 0.2 × 0.35 = 0.623.

[0047] S104: Filter outliers on changes in structural features of adjacent frames based on a dynamic threshold algorithm to generate structural state change data.

[0048] In one implementation, the initial data on deformation displacement, stress fluctuation range, and material property attenuation rate of structural feature changes between adjacent frames are processed in conjunction with a dynamic threshold algorithm to generate structural feature change anomaly filtering information. The dynamic threshold algorithm dynamically adjusts the threshold by analyzing the data distribution in real time. Compared with a fixed threshold, it is more adaptable to data fluctuations. The formula is: in, is the current time mean, is the standard deviation, and k is the dynamic coefficient (adaptive adjustment from 1.5 to 3.0).

[0049] Deformation displacement: Initial data for the Z-axis displacement of a certain automobile chassis beam in adjacent frames (50ms interval) under braking conditions: [0.12, 0.15, 0.18, 0.25, 0.85, 0.22] (unit: mm), with 0.85 being an abnormal spike. Stress fluctuation range: Initial data for stress fluctuations in adjacent frames of an engine bracket: [80, 85, 90, 120, 88] (unit: MPa), with 120 MPa exceeding the normal fluctuation range. Material decay rate: Basic data for the fatigue decay rate of aluminum alloy components in adjacent frames: [0.02, 0.03, 0.025, 0.15, 0.022] (unit: % / h), with 0.15 being an abnormally high decay value. Calculate the average of the current frame data. =0.24mm, standard deviation =0.25mm, the dynamic coefficient k is automatically set to 2.0 based on historical fluctuations, and the threshold value is 0.24+2.0×0.25=0.74mm, filtering out the abnormal displacement of 0.85mm.

[0050] Based on the dynamic threshold algorithm framework, the deformation displacement, stress fluctuation range, and material performance attenuation rate data after outlier filtering are processed. Dynamic threshold adjustment components and data fluctuation analysis scripts are added to the data processing process to limit the outlier fluctuation range and set threshold trigger conditions and feedback mechanisms. The dynamic threshold framework implemented in Python includes: threshold adjustment components, real-time monitoring of data distribution, and updates every 100ms. and The fluctuation analysis script calculates the rate of change between adjacent frames and triggers adaptive threshold adjustment if the rate of change is greater than 30%.

[0051] For the filtered displacement data [0.12, 0.15, 0.18, 0.22], after adding the component, when the new frame displacement is 0.28mm, the rate of change = (0.28-0.22) / 0.22 = 27% < 30%, and the threshold remains at 0.74mm. If the new frame displacement is 0.75mm, the rate of change = (0.75-0.22) / 0.22 = 241% > 30%, and the dynamic coefficient k is automatically increased to 2.5, and the new threshold = 0.25 + 2.5 × 0.24 = 0.85mm. A feedback mechanism is implemented: when three consecutive frames of data trigger the threshold, an early warning signal is output to the production system.

[0052] By adapting the dynamic threshold algorithm to structural feature change data, we generate data on outlier filtering accuracy and real-time performance. Accuracy calibration compares filtering results with manually annotated outliers and calculates the F1 score (the harmonic mean of precision and recall). Real-time calibration measures algorithm processing latency, requiring it to be ≤50ms to meet real-time analysis requirements. Of 100 sets of displacement data containing anomalies, the dynamic threshold algorithm correctly identified 92 with an F1 score of 0.92. The processing latency averaged 42ms, meeting real-time requirements. When equipment vibration increased data noise, calibration adjusted the threshold dynamic coefficient k from 2.0 to 2.8, improving anomaly identification accuracy from 85% to 95%.

[0053] Structural feature change outlier filtering, outlier filtering accuracy, and real-time data are processed to generate structural state change data. Filtered normal data, accuracy indicators (e.g., F1 = 0.95), and real-time indicators (40ms delay) are integrated into structured data. Deformation displacement: Normal fluctuation range [0.12-0.25] mm, outlier filtering rate 95%, processing delay 38ms. Stress fluctuation range: Normal range [80-95] MPa, outlier filtering rate 98%, processing delay 45ms. Material decay rate: Normal range [0.02-0.03] % / h, outlier filtering rate 90%, processing delay 42ms.

[0054] S105 , processing the structural state change data to generate structural health dynamic assessment data and structural abnormality characteristics.

[0055] In one implementation, structural state change data is subjected to time series trend analysis to generate structural state time series parameters. A sliding window Fourier transform (with a window size of 100 frames) is used to extract time series features. This method can analyze the data's components at different frequencies. For example, the Z-axis displacement change data of a vehicle chassis crossbeam under braking conditions is captured at a sampling frequency of 10Hz for 1000 time steps. For example, the displacement values ​​are 0.12mm, 0.15mm, 0.18mm, 0.25mm, 0.22mm, and so on. A sliding window (with a window width of 100) is applied to the data, and the amplitude and phase of the data within each time window are calculated to generate time series parameters containing frequency components. For example, the amplitude of the 10Hz frequency component is 0.05mm.

[0056] Acquire the spatial distribution of structural state change data and vehicle structural operating condition information. Laser scanning is used to obtain the three-dimensional displacement field of the chassis structure with a resolution of 1mm×1mm. For example, in a local area, the displacement at coordinates (100,200) is 0.15mm, (101,200) is 0.16mm, and (102,200) is 0.18mm. Operating condition information is collected through the vehicle's onboard OBD system to obtain vehicle operating conditions, including braking conditions (vehicle speed drops from 60km / h to 0km / h for 5 seconds), acceleration conditions (acceleration of 0.5g for 10 seconds), and cornering conditions (steering wheel angle of 30° for 8 seconds).

[0057] The spatial distribution data is processed to generate the structural state space characteristic parameters. The spatial gradient and curvature are calculated. The spatial gradient reflects the rate of change of displacement in the spatial direction, while the curvature describes the degree of curvature of the displacement field. Taking the coordinate (101, 201) as an example, the calculated x-direction gradient is 0.02 mm / mm and the y-direction gradient is 0.03 mm / mm. The second-order derivatives are 0.01 in the x-direction, 0.02 in the y-direction, and 0.01 in the xy-direction, resulting in a curvature of approximately 0.05.

[0058] The vehicle's structural operating condition data is processed to statistically analyze the distribution of structural features under different operating conditions. This generates baseline values ​​for the correlation between structural state and operating condition, including the structural state fluctuation range during normal operation and the structural state deviation threshold under abnormal operating conditions. The K-means clustering algorithm is used to group the structural features under different operating conditions. The mean and standard deviation of each data group are then calculated to determine the fluctuation range under normal operating conditions and the deviation threshold under abnormal conditions. Under normal braking conditions, the displacement fluctuation range is [0.12, 0.25] mm (mean 0.18 mm, standard deviation 0.04 mm). The deviation threshold for abnormal braking conditions (such as ABS activation) is greater than 0.3 mm (i.e., the mean plus three times the standard deviation).

[0059] The structural state time series parameters, structural state spatial characteristic parameters, and baseline values ​​associated with the structural state and operating conditions are processed to generate a structural state dynamic change curve and structural risk characteristics. The structural health dynamic assessment data and structural anomaly characteristics are composed of the structural state time series parameters, structural state spatial characteristic parameters, baseline values ​​associated with the structural state and operating conditions, structural state dynamic change curve, and structural risk characteristics. Structural anomaly characteristics include fatigue crack risk probability and key component failure warning indicators. The time series parameters and spatial characteristic parameters are normalized and projected onto a two-dimensional plane, with time (seconds) on the horizontal axis and the normalized characteristic values ​​on the vertical axis, to generate the structural state dynamic change curve.

[0060] Risk feature calculation uses the threshold method to identify abnormal situations. If the feature value at a certain moment exceeds the mean plus 2 times the standard deviation, the risk probability is calculated using the standard normal distribution cumulative function.

[0061] The dynamic change curve shows the change of beam displacement over time during braking. The normal range is [0.12, 0.25] mm. At t=3 seconds, the displacement reaches 0.28 mm, exceeding the normal upper limit by 0.03 mm.

[0062] Structural risk characteristics include the following: Fatigue crack risk probability: calculated to be 0.62%, indicating a certain fatigue crack risk. Key component failure warning indicator: The curvature of the abnormal displacement gradient region (101, 201) is 0.05, exceeding the threshold of 0.03, marking the region as a risk point.

[0063] At a certain moment, the chassis crossmember displacement reaches 0.35mm: this displacement exceeds the normal braking range of [0.12, 0.25]mm. The curvature at the corresponding location is 0.08, exceeding the threshold of 0.03. Under normal braking conditions, the displacement exceeds the abnormal deviation threshold of 0.3mm. The calculated fatigue crack risk probability is 0.003%, triggering a fatigue crack risk warning.

[0064] S106 , performing processing based on preset structural safety determination rules and structural abnormality characteristics to generate structural safety determination data.

[0065] In one implementation, the pre-set structural safety assessment rules are constructed based on industry standards (such as ISO 6469-3), material mechanical parameters (such as yield strength and fatigue limit), and automotive structural design specifications, combined with historical failure data (such as the critical stress value for fatigue cracks in a certain vehicle model's chassis). A level 1 warning is triggered when the stress in a key component exceeds 80% of the material's yield strength; an abnormal fluctuation is identified when the structural deformation displacement exceeds 0.3mm within 100ms; and a high-risk status is flagged when the material fatigue index exceeds 0.6.

[0066] Structural anomaly features are extracted and integrated from dynamic structural health assessment data, including: fatigue crack risk probability (e.g., a component's fatigue crack risk probability of 0.62%); key component failure warning indicators (e.g., a curvature value of 0.05 at a certain point exceeding the threshold of 0.03); and abnormal fluctuations in dynamic change curves (e.g., a displacement of 0.28mm during braking exceeding the normal range of [0.12, 0.25]mm). For an aluminum alloy door component, the following anomaly feature set is provided: fatigue crack risk probability: 0.0062 (0.62%); curvature value of a key location: 0.05 (threshold of 0.03); and displacement during braking: 0.28mm (normal upper limit of 0.25mm).

[0067] Using the weighted sum model, the formula is: ,in, is the feature weight (set according to the degree of influence of the feature on structural safety), is the standardized eigenvalue. The weight distribution is as follows: stress anomaly weight =0.4; deformation anomaly weight =0.3; fatigue abnormality weight =0.3.

[0068] Characteristic standardization, stress characteristic: current stress is 70% of yield strength, standardized value =0.7; deformation characteristics: displacement exceeds the normal range by 32%, standardized value =0.8; Fatigue characteristics: fatigue index 0.35, standardized value =0.35. Safety score calculation: Safety level mapping: A score of 0.665 corresponds to "alert level" (out of a maximum score of 1, safety levels are divided into: safety level ≥ 0.8, alert level 0.6-0.8, warning level < 0.6). Dynamic generation of risk level warning thresholds. Threshold determination method: Combine historical data statistical distribution and fault threshold value, the formula is: ,in, =0.6, =1.5, =0.4 (empirical coefficient), is the mean of historical data, is the standard deviation, and the critical value is the safety limit of the material or structure.

[0069] Stress warning threshold: historical normal stress average =100MPa, standard deviation =20MPa; Material yield strength critical value =250MPa; Level 1 warning threshold = MPa. Deformation warning threshold: historical normal displacement average =0.18mm, standard deviation =0.04mm; critical displacement of structural safety = 0.5mm; first-level warning threshold = mm (rounded to 0.37 mm). When the chassis longitudinal beam stress reaches 200 MPa at a certain moment: the rule matches: 200 MPa > 190 MPa (stress warning threshold), triggering an exception; the safety score is updated.

[0070] Stress characteristic normalized value =200 / 250=0.8 (yield strength 250MPa); new score = 0.4 0.8+0.3 0.8+0.3 0.35=0.32+0.24+0.105=0.665to0.765; risk level upgraded: from "alert level" to "warning level" (the score of 0.765 is still greater than 0.6, but if the stress continues to rise to ≥210MPa, the score may drop below 0.6); judgment data updated: safety score: 0.765; risk level: "warning level"; abnormal characteristics: added "stress seriously exceeds the threshold" (200MPavs190MPa).

[0071] S107, processing the safety determination data and real-time analysis characteristic parameters to generate structural fault warning information and production adjustment control instructions.

[0072] In one implementation, safety assessment data is analyzed for risk levels to generate a structural risk assessment. In the chassis production workshop, sensors collect stress, deformation, fatigue, and other data, which are then weighted and calculated to generate a safety score. This score is then mapped to a risk level based on pre-set rules to aid production decision-making. For example, a batch of "urban SUV chassis" was used to focus on the critical structures of the chassis' longitudinal and cross members.

[0073] Stress data: Real-time stress values ​​collected by sensors in the middle of the longitudinal beam are processed and weighted to a standardized value of 0.8 (weighted at 0.4 due to long-term urban road testing, resulting in some stress accumulation but not exceeding limits). Deformation data: Displacement monitoring at the crossbeam connection: standardized at 0.8 (weighted at 0.3 due to deformation slightly exceeding the average during fully loaded vehicle testing). Fatigue data: Based on cyclic loading tests and online monitoring, standardized at 0.35 (weighted at 0.3 due to the slow accumulation of fatigue caused by the application of new materials).

[0074] Safety score calculation, according to the weighted formula shown in the previous example, substitute the data: Safety score = 0.4 0.8+0.3 0.8+0.3 0.35=0.32+0.24+0.105=0.665.

[0075] The preset mapping rules are as follows:

[0076]

[0077] The known safety score is 0.665, which falls in The interval is judged to be at the alert level.

[0078] The actual production linkage is as follows: Monitoring adjustments: On the chassis production line, the sensor acquisition frequency for this batch of products was increased from 1 time / minute to 3 times / minute, focusing on monitoring longitudinal beam stress and crossbeam deformation data. Quality inspection marking: The production management system automatically marks this batch of chassis, and two specialized tests are added to the quality inspection process: "Ultrasonic testing of stress concentration areas" and "3D scanning review of deformation areas." A yellow warning pops up on the workshop screen, indicating "[Batch No.: CX202503] Chassis Risk Level - Alert Level, Continuous Monitoring of Critical Structures Required." The process engineer's workstation receives a pending task, requiring a monitoring and analysis report to be submitted within two hours.

[0079] The system compares the structural anomaly features in the safety assessment data with the preset safety thresholds to generate a structural safety assessment result. The anomaly features in the safety assessment data (such as stress, deformation, and fatigue values) are compared with the preset safety thresholds one by one to determine the anomaly type and severity.

[0080] Stress anomaly comparison, real-time stress value in safety judgment data: 200MPa; preset stress warning threshold: 190MPa. The comparison result exceeds the threshold by 10MPa and is marked as "severely exceeded".

[0081] Deformation anomaly comparison, real-time displacement: 0.28mm; preset deformation threshold: 0.25mm, the comparison result exceeds the threshold by 0.03mm, marked as "slightly exceeded".

[0082] Fatigue comparison, fatigue index: 0.35, preset fatigue threshold: 0.6, the comparison result shows that it does not exceed the threshold and the status is normal.

[0083] The structural safety assessment results are as follows: stress abnormality: seriously exceeds the threshold (200MPa>190MPa); deformation abnormality: slightly exceeds the threshold (0.28mm>0.25mm); fatigue degree: normal (0.35<0.6).

[0084] The structural risk assessment results and structural safety assessment results are processed to generate structural failure warning information. The risk level and anomaly comparison results are combined to generate a warning message containing the risk level, anomaly details, and recommendations. The warning information includes the following: Risk Level Summary: Alert Level (due to severe stress exceedance); Anomaly Feature Details: Stress Concentration Area: Center of the Chassis Longitudinal Beam (Coordinates (50,50)), Predicted Value: 200MPa (Threshold: 190MPa); Abnormal Deformation Area: Cross Beam (101,201), Displacement: 0.28mm (Threshold: 0.25mm); Warning Recommendation: Reduce load or check welding process.

[0085] An example of a complete warning message is as follows: [Structural Failure Warning], Risk Level: Alert; Abnormal Details: Stress in the center of the chassis longitudinal beam severely exceeds the threshold (200 MPa vs. 190 MPa), and local deformation of the crossbeam slightly exceeds the normal range. Recommended Action: Immediately reduce the vehicle load to below half and arrange for a welding process inspection.

[0086] Obtain real-time vehicle structure status information and production management system control configuration information. Real-time vehicle structure status: collected in real time through on-board sensors, such as: current load: 3000kg (half-load condition); real-time stress distribution: 200MPa in the middle of the longitudinal beam, 85MPa in the cross beam;

[0087] Production management system control configuration: preset adjustment strategies, such as: when stress > 190MPa, the production line deceleration program is automatically triggered; when deformation is abnormal, the welding parameter adjustment range: current is reduced by 5-10%. Real-time status: load 3000kg, longitudinal beam stress 200MPa, crossbeam displacement 0.28mm; control configuration: when stress exceeds 190MPa, the production line speed is reduced by 20%; when deformation exceeds 0.25mm, the welding current is reduced by 8%.

[0088] Based on the real-time status of the vehicle structure and structural fault warnings, control configuration information is processed to generate production adjustment control instructions. Specific execution instructions are generated based on the matching rules between warning information and control configurations. Examples of these rules include: if stress exceeds 10-20 MPa, the production line speed is reduced by 20%; if deformation exceeds 0.03-0.05 mm, the welding current is reduced by 8%.

[0089] When the stress exceeds 10MPa (200-190=10), the production line is slowed down by 20% as follows: Control instruction 1: The production line speed is reduced from 10m / min to 8m / min.

[0090] When the deformation exceeds 0.03mm (0.28-0.25=0.03), the welding current is reduced by 8% as follows: Control instruction 2: The welding robot current is adjusted from 200A to 184A.

[0091] Specific production adjustment control instructions include the following:

[0092] 1. The production line speed is reduced to 8m / min (trigger condition: stress exceeds the threshold of 10MPa);

[0093] 2. The welding robot current is adjusted to 184A (trigger condition: deformation exceeds the threshold of 0.03mm);

[0094] 3. The frequency of real-time monitoring of longitudinal beam stress is increased to 200Hz (enhanced monitoring).

[0095] This application is used for structural safety monitoring in the field of high-end automobile manufacturing. Through multi-dimensional data processing and intelligent analysis, it realizes real-time status assessment and fault warning of automobile structure. First, the manufacturing scene data and basic information of the automobile structure are obtained. After preprocessing and feature extraction, feature analysis data such as edge contour feature point set are generated. With the help of the neural network model, the automobile structure is analyzed in real time to obtain parameters such as stress distribution prediction value, deformation trend vector and material fatigue index. The dynamic threshold algorithm is used to filter out abnormal values ​​of structural feature changes in adjacent frames to generate structural state change data. The data is further processed to generate structural health dynamic assessment data and abnormal features, such as fatigue crack risk probability. Combined with preset safety rules and abnormal features, a safety level score and risk warning threshold are generated. Finally, the safety judgment data and real-time analysis parameters are integrated to generate fault warning information and production adjustment instructions.

[0096] By combining neural networks with dynamic threshold algorithms, the system achieves closed-loop management from data collection to production adjustments, improving the real-time and accuracy of vehicle structural analysis. It can be effectively applied to structural safety monitoring in automobile manufacturing, reducing failure risks, and optimizing production processes.

[0097] In one embodiment, Figure 2 As shown, the present application also provides a real-time structural analysis device based on a neural network, comprising:

[0098] Acquisition module 201, used to acquire manufacturing scene data and basic information of automobile structure;

[0099] The processing module 202 is used to preprocess and extract features from the structural features in the manufacturing scene data to generate feature analysis data, including edge contour feature point sets, feature vector probability distributions, abnormal feature filtering results, weight feature parameters, and structural key point feature values; perform real-time analysis of the automobile structure based on a neural network model to generate real-time analysis feature parameters, including stress distribution prediction values, deformation trend vectors, and material fatigue indicators; perform abnormal value filtering on changes in structural features of adjacent frames based on a dynamic threshold algorithm to generate structural state change data, including deformation displacement, stress fluctuation range, and material performance decay rate; process the structural state change data to generate structural health dynamic assessment data and structural abnormality features, including fatigue crack risk probability and key component failure warning indicators; process the data based on preset structural safety judgment rules and structural abnormality features to generate structural safety judgment data, including safety level scores and risk level warning thresholds; process the safety judgment data and real-time analysis feature parameters to generate structural failure warning information and production adjustment control instructions.

[0100] The computer-readable storage medium provided in the above-mentioned embodiments of the present application and the real-time structural analysis method based on a neural network provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0101] Each embodiment in this application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the evaluation of the real-time structural analysis method based on a neural network, the electronic device, the electronic device, and the readable storage medium embodiment, since they are basically similar to the above-mentioned embodiment of the real-time structural analysis method based on a neural network, the description is relatively simple, and the relevant parts can be referred to the partial description of the embodiment of the real-time structural analysis method based on a neural network.

Claims

1. A real-time structural analysis method based on neural network, characterized in that: include: Obtain manufacturing scenario data and basic vehicle structure information; Preprocessing and feature extraction of structural features in the manufacturing scene data to generate feature analysis data, including extracting edge contour feature point sets based on the structural features in the manufacturing scene data; Calculate the edge contour feature point set to generate a feature vector probability distribution; perform outlier filtering on the feature vector probability distribution to generate an abnormal feature filtering result; perform weighted calculation on the filtered feature data to generate weighted feature parameters; perform weighted adjustment on the feature parameters based on the correlation between the feature points and the structure centroid to generate the structural key point feature value, wherein the feature analysis data includes the edge contour feature point set, the feature vector probability distribution, the abnormal feature filtering result, the weighted feature parameter and the structural key point feature value; Real-time analysis of the vehicle structure is performed based on a neural network model to generate real-time analysis characteristic parameters, including processing the historical stress data of the vehicle structure and the current sensor acquisition data based on a stress distribution prediction model to generate a stress distribution prediction value; processing the deformation data of the vehicle structure under different working conditions, analyzing the deformation trend using a time series neural network, and generating a deformation trend vector; processing material fatigue test data and online monitored material performance parameters, and generating a material fatigue index based on a neural network fatigue damage accumulation model; processing the stress distribution prediction value, deformation trend vector, and material fatigue index to generate real-time analysis characteristic parameters, wherein the real-time analysis characteristic parameters include the stress distribution prediction value, deformation trend vector, and material fatigue index; Based on the dynamic threshold algorithm, abnormal value filtering is performed on the structural feature changes of adjacent frames to generate structural state change data, including deformation displacement, stress fluctuation range, and material performance attenuation rate; Processing structural state change data to generate structural health dynamic assessment data and structural abnormality characteristics, including fatigue crack risk probability and key component failure warning indicators; Based on the preset structural safety judgment rules and structural abnormality characteristics, the system generates structural safety judgment data, including safety level scores and risk level warning thresholds; Process safety judgment data and real-time analysis characteristic parameters to generate structural fault warning information and production adjustment control instructions.

2. The method according to claim 1, wherein Based on the dynamic threshold algorithm, abnormal value filtering is performed on the structural feature changes of adjacent frames to generate structural state change data, including: The initial data of deformation displacement of structural feature changes in adjacent frames, the original data of stress fluctuation range, and the basic data of material performance attenuation rate are processed in combination with the dynamic threshold algorithm to generate structural feature change abnormal value filtering information; Based on the dynamic threshold algorithm framework, the deformation displacement, stress fluctuation range, and material performance attenuation rate data after outlier filtering are processed. Dynamic threshold adjustment components and data fluctuation analysis scripts are added to the data processing process to limit the outlier fluctuation range and set threshold trigger conditions and feedback mechanisms. Generate outlier filtering accuracy and real-time data through adaptive calibration of dynamic threshold algorithm and structural feature change data; The structural feature change outlier filtering, outlier filtering accuracy and real-time data are processed to generate structural state change data.

3. The method according to claim 2, wherein Process the structural state change data to generate structural health dynamic assessment data and structural abnormality characteristics, including: Perform time series trend analysis on the structural state change data to generate structural state time series parameters; Obtain spatial distribution data of structural state change data and vehicle structural operating condition information; Process the spatial distribution data to generate the structural state space characteristic parameters; Processing vehicle structural operating condition information, statistically analyzing the distribution of structural characteristics under different operating conditions, and generating baseline values ​​related to structural status and operating conditions, including the structural status fluctuation range during normal vehicle operation and the structural status deviation threshold under abnormal operating conditions; Process the structural state time series parameters, structural state space characteristic parameters and the baseline values ​​associated with the structural state and working conditions to generate the structural state dynamic change curve and structural risk characteristics; Among them, the structural health dynamic assessment data and structural abnormality characteristics are composed of structural state time series parameters, structural state space characteristic parameters, structural state and working condition correlation baseline values, structural state dynamic change curves and structural risk characteristics. Structural abnormality characteristics include fatigue crack risk probability and key component failure warning indicators.

4. The method according to claim 1, wherein Process safety judgment data and real-time analysis characteristic parameters to generate structural fault warning information and production adjustment control instructions, including: Perform risk level analysis on safety judgment data to generate structural risk level assessment results; Compare the abnormal features of the structure in the safety judgment data with the preset safety threshold to generate a structural safety judgment result; Process the structural risk level assessment results and structural safety determination results to generate structural failure warning information; Obtain real-time status information of vehicle structures and control configuration information of production management systems; The control configuration information is processed based on the real-time status of the vehicle structure and structural fault warning information to generate production adjustment control instructions.

5. A real-time structural analysis device based on a neural network, characterized in that: For implementing the method according to claim 1, the apparatus comprises: Acquisition module, used to obtain manufacturing scene data and basic information of automobile structure; The processing module is used to preprocess and extract features of structural features in manufacturing scene data, and generate feature analysis data, including edge contour feature point sets, feature vector probability distributions, abnormal feature filtering results, weight feature parameters, and structural key point feature values; perform real-time analysis of automobile structures based on neural network models, and generate real-time analysis feature parameters, including stress distribution prediction values, deformation trend vectors, and material fatigue indicators; perform abnormal value filtering on changes in structural features of adjacent frames based on a dynamic threshold algorithm, and generate structural state change data, including deformation displacement, stress fluctuation range, and material performance attenuation rate; process structural state change data to generate structural health dynamic assessment data and structural abnormality features, including fatigue crack risk probability and key component failure warning indicators; process based on preset structural safety judgment rules and structural abnormality features to generate structural safety judgment data, including safety level scores and risk level warning thresholds; process safety judgment data and real-time analysis feature parameters to generate structural fault warning information and production adjustment control instructions.

6. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the neural network-based real-time structural analysis method according to any one of claims 1 to 4 by executing the executable instructions.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the neural network-based real-time structural analysis method according to any one of claims 1 to 4 is implemented.

Images