Method for predicting pedestrian protection headform results based on deep learning

Abstract

The present application relates to the technical field of automobile pedestrian protection prediction, and discloses a method for predicting pedestrian protection head type results based on deep learning, comprising the following steps: A1, extracting feature data and constructing a database: extracting feature data and the corresponding HIC simulation value of the feature data from a pedestrian protection head type simulation model to form an HIC database, wherein the feature data comprises head type feature data and vehicle body structure feature data; A2, constructing a deep learning prediction model: training the HIC database using deep learning until the prediction model accuracy reaches the set requirement, and the deep learning is selected as two algorithms of BP neural network and random forest; A3, predicting the head type HIC: after processing the vehicle feature data of the pedestrian protection head type to be predicted, inputting the prediction model meeting the accuracy to obtain the head type HIC prediction value; and A4, making a head score atlas. The method can realize rapid prediction of HIC at all collision positions in the performance evaluation of pedestrian protection head type.

Description

Technical Field

[0001] This invention relates to the field of prediction technology for pedestrian protection in automobiles, and more specifically to a method for predicting pedestrian head shape results based on deep learning. Background Technology

[0002] With the increase in traffic accident statistics, the deepening of biomechanical research, and the improvement of vehicle design capabilities, various regions around the world have successively carried out research on vehicle pedestrian protection standards and testing methods, and countries have successively promulgated vehicle pedestrian protection regulations and star rating systems. In terms of vehicle driving safety performance, pedestrian head shape performance is one of the important evaluation items of the China Insurance Automotive Safety Index (C-IASI) and C-NCAP, characterized by a large test area, high score weight, and high scoring difficulty.

[0003] In the early stages of vehicle development, pedestrian head impact performance evaluation often relies on CAE simulation analysis technology. The Head Injury Index (HIC) measures the degree of head injury to pedestrians caused by a vehicle collision, thereby ensuring that the developed vehicle design meets pedestrian protection regulations, the China Automotive Safety Index, and NCAP star rating requirements. Pedestrian head impact simulation analysis typically requires building a full-vehicle model, followed by mesh generation, condition setting, and calculation submission. For example, a single round of virtual simulation analysis on the full-vehicle model takes approximately 8-9 days, extending the development cycle. However, with the continuous shortening of vehicle development cycles and the increasing number of optimization iterations, there is a need for a faster and more convenient analysis method to provide timely results for pedestrian head impact performance evaluation. Summary of the Invention

[0004] The present invention aims to provide a method for predicting pedestrian head shape results based on deep learning, so as to achieve rapid prediction of HIC at all collision locations in pedestrian head shape performance evaluation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting pedestrian head shape based on deep learning, comprising the following steps:

[0006] A1. Extract feature data and build database: Extract feature data and corresponding HIC simulation values ​​from the pedestrian protection head shape simulation model to form an HIC database. The feature data includes head shape feature data and vehicle body structure feature data.

[0007] A2, Constructing a deep learning prediction model: Train the HIC database using deep learning until the prediction model accuracy reaches the set requirements. The deep learning algorithms selected are BP neural network and random forest.

[0008] A3, Predicted Head Shape HIC: After processing the vehicle feature data of the pedestrian's head shape to be predicted, input it into the prediction model that meets the accuracy requirements to obtain the predicted head shape HIC value.

[0009] A4, create a head score map.

[0010] The beneficial effects of this solution are as follows: 1. By extracting a large amount of model feature data and corresponding HIC simulation values ​​from the pedestrian protection simulation model, a HIC database is formed; deep learning is used to train the HIC database to obtain an effective prediction model; then, pedestrian protection head shape HIC prediction is performed to quickly obtain the predicted HIC value; finally, a score map is generated based on the predicted HIC value. In the context of continuously shortening vehicle development cycles and increasing optimization iterations, a faster and more convenient analysis method is obtained, enabling rapid prediction of HIC at all positions on the hood, shortening the time cost of early development, and contributing to the overall improvement of vehicle development level.

[0011] 2. Unlike other mesh-based finite element algorithms, the prediction results of this method are based on a trained HIC database. The input for analysis is the feature data in the pedestrian protection simulation model and the corresponding HIC results. Then, after inputting new feature data to be predicted, the corresponding HIC prediction value and evaluation score can be directly output. The prediction model only takes 1 second to output the HIC prediction value after receiving new feature data to be predicted.

[0012] 3. In the early stages of development, a pedestrian protection assessment is conducted on the initial design of the vehicle model. If the design shape causes the pedestrian head shape HIC result to exceed the safety limit, it can be determined that the design is unlikely to pass the pedestrian protection assessment. The design can then be modified in a timely manner, avoiding the need to spend a lot of time in the whole vehicle-level simulation model to determine the assessment result. Compared with obtaining the pedestrian protection HIC through simulation, this method can achieve prediction in the early CAD data stage, which greatly shortens the early development cycle and reduces development costs.

[0013] 4. If the obtained HIC result is large, the more prominent data in the HIC prediction value can be used as the key optimization target when evaluating the simulation model. In subsequent simulations, the target should be analyzed in detail. This will not only help to discover defects in a timely manner, but also help to discover factors affecting vehicle safety performance in vehicle design through these prominent data. Based on this, relevant research can be carried out to make corrections and optimizations, thereby improving the R&D level.

[0014] 5. Based on deep learning, the sample feature data is trained to obtain the prediction model. In subsequent use, the feature data of different vehicle models and the corresponding HIC simulation values ​​collected in the simulation model will continue to increase. Deep learning can continuously absorb these new data, continuously correct and update the prediction model, and further improve the prediction accuracy and versatility of this prediction model.

[0015] Preferably, in A1, the head shape feature data includes the mass of the head impactor, the collision velocity, the collision angle, and the collision location; the vehicle body structure feature data includes the length and width of the hood, the thickness of the inner hood panel, the thickness of the outer hood panel, the material parameters of the inner and outer hood panels, the distance between the inner and outer panels at the collision location, and the distance between the inner panel and the hard point at the collision location; and the material parameters of both the inner and outer hood panels include the elastic modulus, yield stress, and ultimate stress.

[0016] The beneficial effects of this approach are: extracting all important data ensures the comprehensiveness of sample data for the next step of deep learning, thereby guaranteeing the accuracy of the simulation results of the prediction model obtained after training.

[0017] Preferably, in A1, the feature data extraction is performed in ANSA software, and the extracted feature data is compiled in a table.

[0018] Preferably, in A2, the accuracy of the prediction models obtained by two different deep learning algorithms is compared, and the best prediction model is selected as the optimal prediction model. At the same time, the HIC prediction value of the selected optimal prediction model must reach 97% or more of the simulation value obtained by the simulation model to be considered that the accuracy of the prediction model has met the set requirements.

[0019] The beneficial effects of this scheme are: when training the prediction model, its accuracy is required to ensure that the ratio of HIC prediction value to simulation value is 97% or higher, so that the obtained prediction model has a certain degree of universality while ensuring prediction accuracy.

[0020] Preferably, in A3, the feature data obtained after feature extraction processing of the new simulation model is the same as that extracted from the simulation result to be predicted using the same method as in A1.

[0021] Preferably, in A4, the predicted head shape score map is created in Excel based on the obtained HIC predicted value.

[0022] The beneficial effects of this scheme are as follows: based on the location of the head collision target point in the prediction model, the data is synchronized to an Excel spreadsheet. Then, according to the HIC prediction scores of different grid points in the prediction model, different colors are used to fill in the corresponding cells in the Excel spreadsheet, resulting in an intuitive prediction head score map.

[0023] Preferably, the test area for pedestrian head shape assessment includes an adult area and a child area, and the adult area and the child area are respectively tested using an adult head shape impactor and a child head shape impactor.

[0024] The beneficial effect of this solution is to ensure the comprehensiveness of head shape feature data collection.

[0025] Preferably, when establishing a simulation model for pedestrian protection head shape assessment, the design criteria should be based on the China Insurance Research Institute's safety index C-IASI or C-NCAP, or foreign standards such as GTR9, ECE127, and E-NCAP pedestrian protection evaluation procedures.

[0026] The beneficial effects of this solution are: to ensure that the predicted HIC value of pedestrian protection head shape can more accurately represent the HIC value obtained from simulation, thereby enabling rapid prediction of pedestrian protection head shape HIC results in the early stages of vehicle development and shortening the development time of new vehicles. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the prediction method in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating the selection of feature data for the simulation model in an embodiment of the present invention;

[0029] Figure 3 This is a partial example illustration of the HIC database in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram illustrating the acquisition of the engine hood width and length in an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram illustrating the distance between the inner and outer panels of the engine hood in an embodiment of the present invention.

[0032] Figure 6 This is a topology diagram of the BP neural network in an embodiment of the present invention;

[0033] Figure 7 This is a schematic diagram of the random forest algorithm in an embodiment of the present invention;

[0034] Figure 8 This is a head score map drawn based on predicted values ​​in an embodiment of the present invention. Detailed Implementation

[0035] The following detailed description illustrates the specific implementation method:

[0036] Example

[0037] The basic implementation examples are as follows: Figure 1-5 As shown, Figure 1The method shown here, based on deep learning, predicts pedestrian head shape (HBC) results. It leverages existing pedestrian head shape performance evaluation technologies using CAE simulation analysis. CAE simulation analysis establishes different simulation models for different vehicle models, then uses these models to simulate HBC simulation values. Model tests on different vehicle models collect datasets of HBC simulation values. From these datasets, feature data and their corresponding HBC simulation values ​​are extracted to construct an HBC database. Deep learning is then used to train the sample data in the HBC database, learning the inherent patterns and hierarchical representations to discover the distributed feature representations. A general prediction model is then built based on this deep learning-based HBC database. By inputting feature data from the new vehicle model to be predicted into the prediction model, the predicted HBC value for that model can be quickly obtained without complex algorithms. Once the predicted HBC value is obtained, the original vehicle design can be modified, or key targets can be identified during subsequent simulation crash tests. This allows for timely detection of the impact of relevant feature data on the HBC results, providing guidance for early vehicle development and design, and helping to accelerate vehicle development.

[0038] Predicting the head shape for pedestrian protection involves the following steps:

[0039] A1. Extract feature data and build database: Extract feature data and corresponding HIC simulation values ​​from the pedestrian protection head shape simulation model to form an HIC database. The feature data includes head shape feature data and vehicle body structure feature data.

[0040] When evaluating vehicle head protection for pedestrians, CAE simulation technology is crucial for assessing the collision protection performance of the hood to reduce early-stage vehicle development costs. This also ensures the collection of a large amount of pedestrian head protection simulation data as a sample. When using CAE simulation analysis for pedestrian head protection collision simulation tests, a corresponding simulation model needs to be established based on the vehicle being evaluated. Then, within the established simulation model, the collision target points for adult or child head shapes are determined according to pedestrian protection testing methods. The collision HIC (Head Indicator Collision Cost) for each target point is calculated sequentially, obtaining the HIC results for all collision locations. Finally, all HIC values ​​are summarized to evaluate whether the vehicle model can effectively protect the pedestrian's head in a collision.

[0041] When calculating the impact damage value, the composite impact acceleration of the head model is obtained by collecting the acceleration in the X, Y, and Z directions of the head impactor. Then, the head damage value HIC at each target point is obtained through integration. The formula for calculating HIC is as follows:

[0042] Among them, (t-4≤15ms)

[0043] In the above formula: t1 is the start time of calculating HIC, t2 is the end time of calculating HIC, t2-t1≤15ms; a is the collision resultant acceleration.

[0044] When evaluating a vehicle's head protection performance for pedestrians using the Hidden Impact Capacity (HIC) indicator, a higher HIC indicates more severe head injuries. Furthermore, in this embodiment, the simulation model for pedestrian head protection assessment is designed based on the China Insurance Research Institute (C-IASI) or C-NCAP safety indices, or international standards such as GTR9, ECE127, and E-NCAP pedestrian protection evaluation procedures. The same head shape, collision speed, and collision angle are used as test parameters for the same type of simulation test to ensure that the HIC simulation value calculated by the simulation model more accurately reflects the head injuries caused by a real vehicle collision, thus providing a certain degree of accuracy assurance for pedestrian head protection performance testing.

[0045] like Figure 2 As shown, when selecting feature data from the pedestrian protection head shape simulation model, although there are many factors that affect the pedestrian protection head shape HIC simulation value results, it is necessary to first divide the engine hood and windshield into grids and then conduct collision tests on each target point when conducting pedestrian protection head shape tests, whether using actual collision physical tests or simulation analysis tests using models. Therefore, the feature data in this embodiment is extracted from two aspects: head shape feature data and vehicle body structure feature data.

[0046] Head shape feature data includes the mass, collision velocity, collision angle, and collision location of the head impactor. The head impactor is a crucial tool for pedestrian protection research. Head impactors are divided into adult head impactors and child head impactors. Based on the vehicle's markings, different head shapes are used in different areas of the hood, thus requiring the extraction of the aforementioned feature data. Furthermore, in this embodiment, the pedestrian protection head shape test area is divided into an adult area and a child area. Collision tests are conducted in the adult area and the child area using adult head impactors and child head impactors respectively, ensuring the comprehensiveness of the head shape feature data collection.

[0047] The vehicle body structural feature data includes the length and width of the hood, the thickness of the inner and outer hood panels, the material parameters of the inner and outer hood panels, the distance between the inner and outer panels at the collision point, the distance between the inner panel and the hard point at the collision point, and the inner hood panel itself. The material parameters for both the inner and outer hood panels include elastic modulus, yield stress, and ultimate stress. The selection of vehicle body structural feature data is based on existing pedestrian head injury simulation datasets, identifying all important data points regarding how the hood would cause head injuries to pedestrians during collision tests. Extracting all important data ensures the comprehensiveness of the sample data for the next step of deep learning, thereby guaranteeing the accuracy of the simulation results obtained from the trained predictive model.

[0048] like Figure 3 As shown, feature data and corresponding HIC simulation values ​​are extracted from a large dataset of pedestrian protection head shape simulations, and then compiled into a HIC database. In this embodiment, the number of feature data collected in the HIC database exceeds 400. Meanwhile, as... Figure 2 As shown, the specific operation for feature data extraction in A1 is as follows: Feature data is extracted using relevant software, thereby completing the batch extraction of all the aforementioned feature data and their corresponding HIC simulation values ​​from the pedestrian protection head shape simulation data, forming an intuitive HIC database. The feature data extraction steps are as follows:

[0049] (1) When building a pedestrian protection head model using ANSA software, a file named MetaList.txt will be automatically generated. This file contains all the location information of the collision target points and the collision type. The collision type includes the impactor mass (adult or child) and the collision speed. Export this file to complete the initial extraction of the mass, collision speed, collision angle and collision position (X,Y,Z) coordinate feature data of the head impactor at different collision target points.

[0050] (2) Based on the feature data, add columns for engine hood length, width, and angle data, and extract the data as follows: Figure 4 As shown, in order to facilitate the extraction of engine hood geometry data for all vehicle models, the engine hood width is defined as the distance between the left and right hinges of the engine hood, and the length is defined as the perpendicular distance from the center of the engine hood latch to the line connecting hinge A and the hinge.

[0051] (3) The thickness of the inner and outer panels of the engine hood is obtained by finding the T1 value in the *section shell corresponding to the inner and outer panels of the engine hood in ANSA software, and the thickness of the inner and outer panels is extracted and added to the feature data.

[0052] (4) The materials of the inner and outer panels of the engine hood are mainly simulated in LS-DYNA using the No. 24 elastoplastic material model. The main material parameters involved in this material model are elastic modulus, yield stress, and ultimate stress. The relevant parameters can also be directly extracted in ANSA software and then added to the feature data.

[0053] (5) Figure 5 As shown, the distances between the inner and outer panels of the engine hood, or between the inner panel of the engine hood and hard points in the engine compartment, can be calculated in the post-processing software of the ANSA software. The software selects the inner and outer panels, a part of the engine hood, and a set of parts consisting of the inner panel and multiple hard points in the engine compartment. The post-processing software automatically calculates the distances from the collision target points on the engine hood to the inner panel and the distances from the inner panel to the hard points in the engine compartment, and then writes a file containing the distance information corresponding to each collision target point. It should be noted that, to more clearly demonstrate the process of feature data acquisition, this embodiment includes... Figure 5 Color information is retained.

[0054] (5) Training deep learning models typically involves dividing the data into two sets: a training set and a test set. The training set is used for iterative approximation of the machine learning model, ensuring that the parameters in the model achieve a good fit between the input and output. The test set is used to test the trained model and verify its accuracy. The above feature data is imported into Python software. Python's built-in pandas module reads the feature data, shuffles it row by row, and outputs 70% of the data as training data and 30% as test data.

[0055] A2, Construct a deep learning prediction model: Train the HIC database using deep learning until the prediction model's accuracy reaches the set requirements.

[0056] When constructing the deep learning prediction model, the HIC data is first cleaned to remove outliers. In this embodiment, a box plot tool is used for this cleaning. Since the number of collision target points in the pedestrian head-shaped simulation model is in the hundreds, and the simulation calculation method for each grid point is basically the same, the model suffers from high repetition and large computational load. Therefore, deep learning is used to train the model on the sample feature data in the HIC database. Deep learning learns the inherent patterns and representation levels of the sample data to discover the distributed feature representation of the data. Based on the trained feature data with clearly defined inherent patterns, a general prediction model is built. When predicting pedestrian head-shaped results for a new vehicle model, the aforementioned feature data of that vehicle model is input into the prediction model, and the model can quickly provide the corresponding HIC prediction value, thus achieving second-level, high-precision intelligent prediction of pedestrian head-shaped HIC results.

[0057] In this embodiment, deep learning is selected using two algorithms: backpropagation neural network and random forest, to establish a deep learning prediction model.

[0058] Backpropagation (BP) neural networks are multilayer feedforward neural networks trained using an error backpropagation algorithm. BP neural networks can solve complex nonlinear problems, and trained and validated BP neural networks offer high prediction accuracy and short processing time. BP neural networks possess the ability to continuously learn iteratively, acquiring nonlinear mapping relationships between different data to improve their robustness. The learning process consists of two parts: forward propagation of the signal and backward propagation of the error. Forward propagation involves neurons in the input layer receiving various information from the outside world and passing it to neurons in the intermediate hidden layers. These neurons process and transform the received information according to requirements. Backpropagation occurs when the error between the actual output and the ideal output exceeds the expectation. Starting from the output layer, the error is corrected using gradient descent to adjust the weights of each layer, propagating sequentially to the hidden layers and then the input layer. Through continuous forward and backward propagation, the weights of each layer are constantly adjusted. Training ends when the output error decreases to the expected level or the pre-set number of learning iterations.

[0059] The topology of a BP neural network is as follows: Figure 6 As shown. Where X1, X2, ... X n These are the input values ​​of the BP neural network, namely, head shape and vehicle body structure feature data. ij and w jk Y represents the weights of the BP neural network. i This is the output value of the BP neural network, namely the HIC value of head injury. When the number of input nodes is n and the number of output nodes is m, the BP neural network reflects the functional mapping relationship from n independent variables to m dependent variables.

[0060] Another important factor is the selection of the number of hidden layer nodes, as this significantly impacts the performance of a backpropagation (BP) network. Currently, there is no ideal analytical formula to determine the optimal number of neurons; the common approach is to use an empirical formula to provide an estimate. The formula is as follows:

[0061]

[0062] In the formula: l is the number of hidden layer nodes, m is the number of output layer nodes, n is the number of input layer nodes, and a is a constant between 0 and 10. Through data preprocessing, the prediction model has 14 input parameters, 1 output parameter, and 6 hidden layer nodes. Therefore, the neural network structure is set to 4-6-1, meaning the input layer has 4 nodes, the hidden layer has 6 nodes, and the output layer has 1 node.

[0063] By importing feature data and training it with a backpropagation (BP) neural network, the optimal parameters are found, and the best model is obtained, thus constructing the HIC prediction model. The continuous and complete feature data to be predicted is then input into the trained HIC prediction model for prediction. The average of the prediction results from all decision trees is the final predicted HIC result.

[0064] The Random Forest algorithm boasts advantages such as high accuracy, the ability to handle large-scale datasets, no need for feature normalization or handling missing values, and the ability to assess feature importance. Based on decision trees, the Random Forest algorithm randomly selects K new datasets with replacement from the original training dataset, generating K decision trees to form a random forest. The final prediction result is the mean of the predictions from all decision trees. The basic flow of the model is shown in the figure, and the basic steps are as follows:

[0065] Head shape feature data and vehicle body structure feature data are used as input parameters, and pedestrian head injury index values ​​are used as output parameters. These are input into a random forest model for model training. During training, the feature data is divided into two parts: 70% of the data is used to train the random forest model, and 30% of the data is used to validate the model's accuracy.

[0066] like Figure 7 As shown, when training feature data using random forests, the hyperparameters of the model are adjusted, including: depth (here, the depth of each decision tree in the forest), number of decision trees, percentage of variables used in each decision tree, and minimum number of samples for leaf splits.

[0067] (1) Apply the bootstrap method to randomly sample N datasets with replacement from the original training set S to generate N decision trees. Start training the model, using head shape feature data, body structure feature data and corresponding HIC simulation values ​​as the original training set S. Apply the bootstrap method to randomly sample datasets with replacement from the original dataset S to randomly generate N training subsets DN, and build the corresponding N decision trees. (2) The decision tree adopts the CART decision tree. Each time a branch grows, m features (m≤M) are randomly selected from M feature attributes. The index for measuring the quality of the branch is the mean squared error (MSE). Use the head shape feature data and body structure feature data in the input variables as the original features. Randomly sample M features from the original features as the split feature set of the node. Select the optimal feature based on the mean squared error so that the decision tree makes judgments based on the optimal feature, so that the decision tree grows continuously and gets closer to the optimal result. No pruning is performed in the middle process. (3) Select the optimal feature based on the mean squared error to maximize branch growth. No pruning is performed in the middle process. The random forest regression model is used for training. Different types of hyperparameters are arranged into a grid. Each set of hyperparameters in the grid is cross-validated multiple times (e.g., more than 5 times). That is, the grid search and 5-fold cross-validation are used to optimize the model, find the optimal parameters, and obtain the best model, that is, to construct the HIC prediction model. (4) The average of the prediction results of all decision trees is the final prediction result. The continuous and complete feature data to be predicted is input into the trained HIC prediction model for prediction. The average of the prediction results of all decision trees is the final predicted HIC result.

[0068] Furthermore, the prediction accuracy of the prediction model refers to the ratio of the HIC predicted values ​​on the validation set output by the prediction model to the HIC simulated values ​​output by the simulation model. If the ratio reaches 97% or higher, the established prediction model is considered effective and meets the set requirements, and the next step can be carried out. Based on the 97% accuracy, the prediction model accuracy obtained by the BP neural network algorithm and the random forest algorithm is compared, and the prediction model with the highest accuracy is selected as the final prediction model. If the prediction model accuracy obtained by both algorithms reaches 97% or higher, it is necessary to return to A1, supplement the feature data extraction again, and perform deep learning on the HIC database with the newly added feature data to retrain the sample feature data until the prediction model meets the set requirements, and then select the target model with higher accuracy again.

[0069] Therefore, this embodiment uses the feature data of pedestrian protection head shapes from multiple vehicle models and the corresponding HIC simulation values ​​as new learning samples, enabling the prediction model to be continuously corrected and updated, training the prediction model until the prediction accuracy reaches the set requirements. Simultaneously, to verify the effectiveness of the prediction model, in this embodiment, 70% of the feature data is used for model training, and the remaining 30% is used for verification, comparing the simulated HIC values ​​with the predicted values ​​until the predicted value reaches 95% of the simulated value.

[0070] A3, Predicted Head Shape HIC: After processing the vehicle feature data of the pedestrian's head shape to be predicted, input it into a prediction model that meets the required accuracy to obtain the predicted head shape HIC value.

[0071] After deep learning training, the established prediction model can effectively provide the pedestrian protection HIC (Hazard Indicator) results for the new vehicle model to be predicted. First, the data from the vehicle simulation model to be predicted is extracted sequentially according to the method in A1. Then, the feature data is input into the prediction model for analysis and processing, and finally, the predicted HIC results are output. Unlike other mesh-based finite element algorithms, this method's prediction results are based on a pre-trained HIC database. The input for analysis is the feature data from the pedestrian protection simulation model and its corresponding HIC results. Therefore, after inputting new feature data to be predicted, the model can directly output the predicted HIC value and evaluation score based on the constructed deep learning model. In this embodiment, the prediction model only needs 1 second to output the predicted HIC value after receiving new feature data. Compared to the existing simulation model, which requires 8-9 days to run a single data verification system, this method achieves a fast and convenient prediction objective.

[0072] In the early stages of vehicle development, feature data from the initial design of the vehicle model can be imported into a predictive model for rapid HIC (Hazard Analysis and Criterion) prediction. If the HIC result is large, it can be determined that the initial design of the vehicle model is unlikely to pass the pedestrian protection assessment, and the design can be modified in a timely manner. This avoids spending a lot of time in the whole vehicle-level simulation model to determine the assessment result, which greatly shortens the early development cycle and reduces development costs.

[0073] A4, create a head score map.

[0074] After obtaining the predicted Hierarchical Indicators (HICs) of each target point on the engine hood in the prediction model, the predicted head HIC results are plotted on an Excel spreadsheet to obtain the predicted head score map.

[0075] like Figure 8As shown, the predicted values ​​obtained from deep learning are displayed in an Excel format as HIC scores. The target points for head collisions in the prediction model are synchronized to the Excel spreadsheet. Based on the HIC prediction scores of different grid points in the prediction model, different colors are used to fill in the corresponding cells in the Excel spreadsheet, thus obtaining an intuitive predicted head score map. It should be noted that, to more clearly demonstrate the obtained head score map, this embodiment includes... Figure 8 Color information is retained.

[0076] In this prediction method, a large amount of model feature data and corresponding HIC simulation values ​​are extracted from the pedestrian protection simulation model to form an HIC database. Deep learning is then used to train the HIC database to obtain an effective prediction model. Next, pedestrian protection head shape HIC prediction is performed to quickly obtain the predicted HIC value. Finally, a score map is generated based on the predicted HIC value. Therefore, in the context of continuously shortening vehicle development cycles and increasing optimization iterations, a faster and more convenient analysis method is obtained, enabling rapid HIC prediction, reducing early development time costs, and contributing to the overall improvement of vehicle development level. Furthermore, the prediction method mentioned in this embodiment is also applicable to obtaining all the above feature parameters and corresponding HIC values ​​through experimental methods, and then training the model using experimental data to predict pedestrian protection head shape HIC values.

[0077] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting pedestrian head shape for protection based on deep learning, characterized in that: Includes the following steps: A1. Extract feature data and build database: Extract feature data and corresponding HIC simulation values ​​from the pedestrian protection head shape simulation model to form an HIC database. The feature data includes head shape feature data and vehicle body structure feature data. A2, Constructing a deep learning prediction model: Train the HIC database using deep learning until the prediction model accuracy reaches the set requirements. Select two algorithms, BP neural network and random forest, to build a deep learning prediction model. By comparing the accuracy of the prediction models obtained by the BP neural network algorithm and the random forest algorithm, select the prediction model with the highest accuracy as the final prediction model. A3, Predicted Head Shape HIC: After processing the vehicle feature data of the pedestrian's head shape to be predicted, input it into the prediction model that meets the accuracy requirements to obtain the predicted head shape HIC value. A4, create a head score map; In A1, feature data extraction is performed in ANSA software, and the extracted feature data is compiled in a table. The feature data extraction steps are as follows: Extract all collision target point location information and collision type from the MetaList.txt file automatically generated when building the pedestrian protection head model using ANSA software. This completes the initial extraction of feature data for the mass, collision velocity, collision angle, and collision position (X,Y,Z) coordinates of the head impactor at different collision target points. Based on the feature data, the engine hood length and width are further added. The T1 values ​​in the *sectionshell corresponding to the inner and outer parts of the engine hood are found in ANSA software to obtain the thickness of the inner and outer parts, which are then added to the feature data. The material parameters of the inner and outer panels of the engine hood are directly extracted in ANSA software and added to the feature data; In the ANSA software's post-processing software, select the inner and outer panels of the engine hood, as well as a set of parts consisting of the inner panel of the engine hood and multiple hard points in the engine compartment. The post-processing software automatically calculates the distance between the collision target point on the engine hood and the inner panel, and the distance between the inner panel and the hard points in the engine compartment. Then, it writes a file containing the distance information corresponding to each collision target point, thus obtaining the distance information corresponding to each collision target point.

2. The method for predicting pedestrian protection head shape results based on deep learning according to claim 1, characterized in that: In A2, the accuracy of the prediction models obtained by two different deep learning algorithms is compared, and the best prediction model is selected as the optimal prediction model. At the same time, the HIC prediction value of the selected optimal prediction model must reach 97% or more of the simulated value obtained by the simulation model to be considered to meet the set requirements for the accuracy of the prediction model.

3. The method for predicting pedestrian protection head shape results based on deep learning according to claim 2, characterized in that: In A3, the feature data obtained after feature extraction processing of the new simulation model is the same as that extracted from the simulation results to be predicted using the same method as in A1.

4. The method for predicting pedestrian protection head shape results based on deep learning according to claim 3, characterized in that: In A4, based on the obtained HIC prediction values, a predicted head shape score map is created in Excel.

5. The method for predicting pedestrian protection head shape results based on deep learning according to any one of claims 1-4, characterized in that: The test areas for pedestrian head shape assessment include an adult area and a child area, with adult head shape impactors and child head shape impactors used for crash tests in the adult area and child area, respectively.

Images