A method, device, and storage medium for predicting a head injury criterion of an occupant in a collision

Abstract

The application discloses a kind of prediction method, device and storage medium of passenger collision head injury grade, relating to injury assessment technical field, which comprises the following steps: first, source domain data and target domain data are obtained;Then the two types of data are preprocessed;Subsequently, the preprocessed source domain data is input into the CNN-LSTM hybrid model for pre-training to obtain a pre-training model;Then, the parameters of the CNN part of the pre-training model are frozen, and the LSTM part and the fully connected layer are fine-tuned using the preprocessed target domain data to obtain a transfer learning model for the target posture;Finally, a ride angle selection model is selected: if greater than the angle threshold, the transfer learning model is used to output the injury grade under the target posture;If less than or equal to the angle threshold, the pre-training model is used to output the injury grade under the standard posture.The method can effectively improve the accuracy of passenger injury prediction during the collision process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of injury assessment, and in particular to a method and device for predicting the head injury level of an occupant in a collision, equipment and a storage medium. BACKGROUND

[0002] With the rapid development of intelligent and electric technologies of automobiles, and the continuous improvement of user demand for ride comfort, seat configurations with large-angle reclining or even complete lying functions have gradually become important selling points in high-end passenger cars and autonomous vehicles. However, the significant change in seat posture not only improves the comfort experience, but also brings new challenges to vehicle collision safety design. The traditional occupant restraint system and safety performance evaluation are mainly based on the standard sitting posture. Once the seat is in a large reclining angle, the interaction relationship between the occupant and the restraint system such as the seat belt and airbag will change fundamentally, and the effectiveness of the original safety strategy is facing severe tests.

[0003] At present, the mainstream method for predicting occupant collision injury mainly relies on finite element models, multi-rigid-body dynamics models and physical collision test databases based on standard sitting posture. These methods are relatively mature in the conventional posture and can accurately evaluate the dummy dynamics response and injury risk. In addition, in recent years, some research has tried to introduce machine learning methods to construct injury prediction models based on simulation or test data to improve the modeling efficiency and generalization ability.

[0004] However, when facing the scene of large-angle seat posture, the existing method has low accuracy in predicting the injury of the occupant during the collision process. SUMMARY

[0005] The present application provides a method and device for predicting the head injury level of an occupant in a collision, equipment and a storage medium, which can improve the accuracy of predicting the injury of the occupant during the collision process when facing the scene of large-angle seat posture.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a method for predicting the head injury level of an occupant in a collision, comprising:

[0008] Obtaining source domain data and target domain data, the source domain data being occupant collision data under a standard posture, the target domain data being occupant collision data under a target posture, the sitting angle corresponding to the target posture being greater than an angle threshold, and the sample number of the target domain data being less than the sample number of the source domain data;

[0009] Preprocessing the source domain data and the target domain data;

[0010] The preprocessed source domain data is input into a convolutional neural network-long short-term memory network (CNN-LSTM) hybrid model for pre-training, to obtain a pre-training model;

[0011] Parameters of a CNN part of the pre-training model are frozen, and the pre-training model is fine-tuned using preprocessed target domain data, to obtain a transfer learning model for a target posture;

[0012] If the riding angle is greater than the angle threshold, parameters corresponding to the target posture are input into the transfer learning model, and a head injury level of an occupant in the target posture is output;

[0013] If the riding angle is less than or equal to the angle threshold, parameters corresponding to the standard posture are input into the pre-training model, and a head injury level of an occupant in the standard posture is output.

[0014] Optionally, the pre-processing of the source domain data and the target domain data comprises:

[0015] The source domain data and the target domain data are normalized respectively;

[0016] The normalized source domain data and the target domain data are processed for dimension reduction.

[0017] Optionally, the processing for dimension reduction of the normalized source domain data and the target domain data comprises:

[0018] The normalized source domain data and the target domain data are processed for dimension reduction by a Gram angle field algorithm.

[0019] Optionally, the freezing of the parameters of the CNN part of the pre-training model and the fine-tuning of the LSTM part and the fully connected layer of the pre-training model using the preprocessed target domain data comprises:

[0020] The weight parameters of the convolutional layer and the pooling layer of the CNN part of the pre-training model are kept unchanged;

[0021] The preprocessed target domain data is input into the pre-training model, to obtain a prediction result;

[0022] Based on a loss between the prediction result and a real label, the weight parameters of the LSTM part and the fully connected layer are updated by a back propagation algorithm.

[0023] Optionally, the method further comprises:

[0024] When a prediction accuracy rate of the injury level of the transfer learning model on a validation set is greater than or equal to an accuracy threshold, the fine-tuning training is stopped, and the transfer learning model for the target posture is obtained.

[0025] Optionally, the method further comprises:

[0026] During the fine-tuning process, monitor the inter-domain differences between the feature distribution of the target domain data and the feature distribution of the source domain data;

[0027] When the inter-domain difference is greater than the difference threshold, the parameters of the deep convolutional layers in the CNN part of the pre-trained model are unfrozen.

[0028] Based on the inter-domain difference, the learning rates of the thawed deep convolutional layer parameters, some LSTM parameters, and fully connected layer parameters are adjusted respectively.

[0029] Optionally, the occupant collision data includes the spatial coordinates of multiple fixed points of the seat belt, the spatial coordinates of point H of the dummy, the collision acceleration waveform, and the labeled occupant head injury level.

[0030] Secondly, this application provides a device for predicting the degree of head injury in a collision, comprising:

[0031] The acquisition module is used to acquire source domain data and target domain data. The source domain data is occupant collision data under standard posture, and the target domain data is occupant collision data under target posture. The riding angle corresponding to the target posture is greater than the angle threshold, and the number of samples in the target domain data is less than the number of samples in the source domain data.

[0032] The data processing module is used to preprocess the source domain data and the target domain data; input the preprocessed source domain data into the CNN-LSTM hybrid model for pre-training to obtain a pre-trained model; freeze the parameters of the CNN part in the pre-trained model; and use the preprocessed target domain data to fine-tune the LSTM part and the fully connected layer in the pre-trained model to obtain a transfer learning model of the target pose.

[0033] The prediction module, if the riding angle is greater than the angle threshold, inputs the parameters corresponding to the target posture into the transfer learning model and outputs the occupant's head injury level under the target posture.

[0034] If the seating angle is less than or equal to the angle threshold, the parameters corresponding to the standard posture are input into the pre-trained model, and the occupant's head injury level under the standard posture is output.

[0035] Thirdly, this application provides a computing device, including a memory and a processor;

[0036] The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0037] In a fourth aspect, the present application provides a computer readable storage medium for storing a computer program for executing the method according to any one of the first aspect.

[0038] From the above technical solutions, the present application has at least the following beneficial effects:

[0039] In the present application, in view of the problems of lack of large-angle posture collision test data and invalidation of traditional prediction models based on standard sitting posture due to dramatic changes in the interaction between occupants and restraint systems, the present application fully reuses the damage correlation knowledge contained in the rich source domain data under the standard posture by means of the transfer learning strategy, and fine-tunes the model with a small amount of target domain data, greatly reducing the dependence on expensive physical tests or high-fidelity simulation data for large-angle postures, and improving the prediction accuracy of occupant head injury level under large-angle postures on the premise of controllable data cost.

[0040] Further, the present application cooperatively designs the pre-training model and the transfer learning model, and adaptively matches the corresponding prediction model for different seat sitting angles. When the sitting angle does not exceed the threshold, the pre-training model trained by sufficient source domain data is directly used to ensure the prediction accuracy under the standard posture. When the sitting angle exceeds the threshold, the transfer learning model adapted to the large-angle scene is enabled, realizing full-scene coverage from the regular sitting posture to the large-angle lying posture, and meeting the safety evaluation needs of the diversified seat configuration of intelligent and electric vehicles.

[0041] Further, the present application introduces an inter-domain difference monitoring mechanism. When the feature distribution difference between the target domain and the source domain is large, the CNN deep layer parameters are dynamically unfrozen and the learning rate is adjusted, which not only avoids the model fine-tuning shock caused by too large inter-domain difference, but also fully excavates the learning ability of the model to specific features of large-angle postures. At the same time, the accuracy threshold is used to control the fine-tuning termination time, which improves the model training efficiency while ensuring the prediction accuracy.

[0042] Further, compared with the traditional method which relies on a large number of physical tests or complex dynamic modeling, the present application based on the transfer learning framework of the CNN-LSTM hybrid model greatly shortens the period of safety performance evaluation of large-angle seats, reduces the research and development cost, and provides an efficient and accurate safety evaluation tool for the large-angle seat configuration of high-end passenger cars and autonomous vehicles, helping to improve the riding comfort of related vehicle models while building a strong collision safety line.

[0043] It should be understood that the description of technical features, technical solutions, advantages or similar language in this application does not imply that all features and advantages can be achieved in any single embodiment. On the contrary, it can be understood that the description of features or advantages means that the specific technical features, technical solutions or advantages are included in at least one embodiment. Therefore, the description of technical features, technical solutions or advantages in this specification does not necessarily refer to the same embodiment. Further, the technical features, technical solutions and advantages described in this embodiment can be combined in any appropriate manner. Those skilled in the art will understand that the embodiments can be implemented without one or more specific technical features, technical solutions or advantages of the specific embodiments. In other embodiments, additional technical features and advantages can be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A flowchart of a method for predicting the head injury level of an occupant in a collision provided in an embodiment of the application;

[0045] Figure 2 A CNN network structure provided in an embodiment of the application;

[0046] Figure 3 An LSTM network structure provided in an embodiment of the application;

[0047] Figure 4 A schematic diagram of a device for predicting the head injury level of an occupant in a collision provided in an embodiment of the application;

[0048] Figure 5 A schematic diagram of a computing device provided in an embodiment of the application. DETAILED DESCRIPTION

[0049] The terms "first", "second", and "third" and the like in the specification and the drawings of this application are used to distinguish different objects, not to limit a specific order.

[0050] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration, in no way limiting. The embodiments or designs described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as more preferred or advantageous than other embodiments or designs. Rather, the use of "exemplary" or "for example" is intended to present the relevant concept in a specific manner.

[0051] For the sake of clear and concise description of the following embodiments, first give a brief introduction of the related art:

[0052] Transfer learning is a machine learning strategy for cross-domain knowledge reuse, aiming to transfer the model knowledge trained from a data-rich source domain (standard posture) to a data-scarce target domain (large-angle posture), to achieve high-precision prediction of the target domain task through a small amount of target domain data fine-tuning, and to avoid the underfitting problem caused by model training from zero.

[0053] With the rapid development of automobile intelligence and electrification, large-angle seats have become an important configuration of high-end passenger cars and autonomous vehicles. However, the existing occupant crash injury prediction technology is difficult to adapt to this scenario. The main technical problem is that the prediction accuracy of occupant head injury in a large-angle posture is low, which cannot meet the safety evaluation requirements. Traditional finite element models, multi-rigid-body dynamics models, and conventional machine learning models are all based on standard postures, and their core assumption is that the interaction between the occupant and the restraint system is fixed. However, in a large-angle posture, the occupant's body tilts backward, causing the dummy H-point to move upward, the distance between the head and the airbag to shorten, and the relative position of the seatbelt anchor point to change. This fundamentally changes the restraint force transmission path of the restraint system and the dynamic response pattern of the occupant during the crash, causing the core assumption of traditional models to fail, which directly leads to a sharp decline in prediction accuracy.

[0054] The main causes of this technical problem also include the objective constraints of target domain data scarcity and the technical challenges of cross-domain feature distribution differences. Physical crash tests in a large-angle posture have high equipment requirements, high material consumption, and high difficulty in scene replication. The cost of a single test is high, and it is difficult to cover all large-angle working conditions and occupant body types, resulting in a much smaller number of target domain data samples than source domain data. The label is scarce and cannot support the full training of traditional deep learning models, forming a vicious cycle of data scarcity, model underfitting, and inaccurate prediction. At the same time, there are natural differences between standard postures and large-angle postures in static spatial relationships (such as the initial distance between the head and the airbag) and dynamic temporal changes (such as the occupant's forward leaning speed and the timing of the acceleration peak). This causes the feature distribution of the source domain and the target domain data to be unable to be directly aligned, and the source domain trained model applied to the target domain will further exacerbate the prediction error due to insufficient generalization ability, ultimately making the crash safety evaluation of large-angle seats a key technical problem that needs to be solved in the process of intelligent upgrading of the automotive industry.

[0055] Therefore, the embodiments of the present application provide a method for predicting the injury level of the occupant's head in a crash. The method can be executed by a processing device. The processing device can be a terminal or a server. The terminal includes but is not limited to a vehicle terminal, a smartphone, a tablet computer, a notebook computer, a personal digital assistant, or a smart wearable device, etc. The server can be a cloud server, such as a central server in a central cloud computing cluster or an edge server in an edge cloud computing cluster. Of course, the server can also be a server in a local data center. The local data center refers to a data center directly controlled by the user.

[0056] To solve the problem of low prediction accuracy of occupant collision injury in large-angle seat posture, first, the standard posture rich source domain data is used to learn the spatial feature and time sequence feature correlation law in the collision process through the CNN-LSTM hybrid model, and a pre-trained model with stable performance is constructed. Then, based on the transfer learning strategy, the CNN part parameters with general spatial feature extraction ability in the model are frozen, and only the LSTM part and the full connection layer good at capturing dynamic time sequence features are fine-tuned with a small amount of target domain data to quickly adapt to the specificity of the large-angle posture scene. At the same time, a dynamic monitoring mechanism of inter-domain difference is introduced, and the parameter freezing / thawing state and learning rate are flexibly adjusted according to the feature distribution difference between the source domain and the target domain to ensure the adaptability and prediction stability of the model. Finally, by judging the relationship between the seat sitting angle and the threshold value, the pre-trained model (standard posture) or the transfer learning model (large-angle posture) is adaptively called to realize the prediction of the occupant head injury level in all scenes.

[0057] In order to make the technical solutions of the present application clearer and easier to understand, the following will introduce a method for predicting the head injury level of an occupant in a collision provided by an embodiment of the present application in conjunction with the accompanying drawings. As shown in the figure, it is a flow chart of a method for predicting the head injury level of an occupant in a collision provided by an embodiment of the present application. The method comprises: Figure 1

[0058] S201, the processing device acquires source domain data and target domain data, the source domain data is occupant collision data in a standard posture, the target domain data is occupant collision data in a target posture, the sitting angle corresponding to the target posture is greater than an angle threshold, and the sample quantity of the target domain data is less than that of the source domain data.

[0059] The source domain refers to the basic reference field (standard posture collision scene) for model training. The source domain data is the data collected based on this field for model pre-training. The data type is complete, the sample quantity is sufficient, and it is the basic data source for the model to learn the correlation law of collision injury.

[0060] The target domain refers to the application field (large-angle posture collision scene) that the model finally needs to adapt to. The target domain data is the supplementary data collected based on this field for model fine-tuning. The data type is consistent with that of the source domain data, but the sample quantity is significantly less. The role is to adapt the model to the specificity of the large-angle posture scene.

[0061] The standard posture is the traditional reference sitting posture for automobile collision safety design and test, which refers to the included angle between the seat backrest and the cushion being 25-35°. At this time, the occupant's body is upright, and the spatial position relationship with the restraint system such as the safety belt and the airbag is relatively fixed. It is a conventional reference posture for industry safety performance evaluation.

[0062] ​The target posture is a large-angle riding posture that needs to be adapted in the present application, i.e., a posture in which the angle between the seat back and the cushion is greater than the angle threshold, covering a state of a large inclination to a near lying state, commonly seen in high-end passenger cars and autonomous vehicles, and characterized by a significant change in the spatial position relationship between the occupant and the restraint system.

[0063] The riding angle is the angle between the seat back and the cushion, which is a quantitative index for distinguishing between the standard posture and the target posture, and directly determines the riding posture of the occupant and the interaction relationship with the restraint system.

[0064] The angle threshold is a critical angle value for dividing the standard posture and the target posture, which is set to 45°. When the riding angle is greater than the angle threshold, it is determined to be a large-angle scenario that needs to be adapted; when the riding angle is less than or equal to the angle threshold, it is determined to be a traditional standard scenario.

[0065] The occupant collision data is the physical quantity data related to the occupant injury in the collision scenario, including the spatial coordinates of the multiple fixed points of the safety belt, the spatial coordinates of the dummy H point, the collision acceleration waveform, and the labeled occupant head injury level, which is the data carrier for model training and prediction.

[0066] The spatial coordinates of the multiple fixed points of the safety belt are used as the representation of the spatial posture of the restraint system, which directly defines the restraint position and the force transmission path of the safety belt to the occupant during the collision, and its dynamic change reflects the adaptation relationship between the restraint system and the occupant's body.

[0067] The spatial coordinates of the dummy H point are used as the reference for the occupant's body posture, and its dynamic change directly reflects the overall displacement, posture deflection of the occupant during the collision, and the relative position relationship with the vehicle structure, which is the basis for judging whether the occupant is facing a secondary collision risk.

[0068] The collision acceleration waveform is used as a quantitative index of the collision impact intensity, and its peak value, rise time, and duration are directly related to the stress on the key parts of the occupant.

[0069] The labeled occupant head injury level is used as the supervision label for model training, which is a standardized and quantitative representation of the injury results corresponding to the aforementioned physical quantity data.

[0070] The processing device uniformly obtains two types of data sources: the source domain data corresponds to the standard posture collision scenario with the angle between the seat back and the cushion ≤ the angle threshold, which is a mature data source in the industry, with sufficient samples and complete data types (including occupant collision-related physical quantities and injury level labels); the target domain data corresponds to the large-angle target posture collision scenario with the riding angle > the angle threshold, and the data types are consistent with the source domain. Due to the high cost and difficulty of large-angle collision tests, the sample size is significantly less than that of the source domain. This setting is consistent with the actual engineering limitations and also forms the logical basis for introducing the transfer learning strategy in the present application.

[0071] S202. The processing equipment preprocesses the source domain data and the target domain data.

[0072] Specifically, the processing device normalizes the source domain data and the target domain data respectively; and performs dimensionality reduction on the normalized source domain data and the target domain data.

[0073] First, normalization is performed on the source and target domain data respectively. By unifying the units, numerical differences between different features are eliminated. For example, large values ​​of acceleration are avoided from masking small-range changes in coordinates. On this basis, the Gram angle field algorithm is used to reduce the dimensionality of the normalized source and target domain data, simplifying the data structure and preserving the characteristic patterns of the collision process (such as acceleration peaks and coordinate change trends). Finally, a low-dimensional and efficient feature representation adapted to the input of the CNN-LSTM model is obtained, providing a high-quality data foundation for subsequent model pre-training and cross-domain fine-tuning.

[0074] The calculation expression for normalization is:

[0075]

[0076] in, Indicates the first The feature parameters at time step Normalized value at time, Indicates the first The feature parameters at time step The original values ​​at that moment, such as the coordinates of the seatbelt anchor point and the collision acceleration value at a certain instant. Indicates the first The minimum value of each feature parameter at all time steps. Indicates the first The maximum value of each feature parameter across all time steps.

[0077] The purpose of dimensionality reduction is to transform a one-dimensional time-series data sequence X of length n into an n×n two-dimensional image feature matrix G. Assume the normalized one-dimensional time-series data sequence... ,in, Let X represent the total number of time steps in the one-dimensional time series data sequence, and let X represent the one-dimensional time series data sequence to be preprocessed. This represents a one-dimensional time-series data sequence after normalization. Indicates the first The feature parameters at time step The normalized value at time.

[0078] One-dimensional time series data sequence after normalization Each time step in the value of the angle angle encoding is performed to obtain the corresponding angle value .

[0079]

[0080] wherein, denotes the angle value after angle encoding of is used for subsequent construction of two-dimensional image features.

[0081] Using the angle value of each time step obtained in the previous step, the Gram matrix G is calculated, and the elements in the Gram matrix G are calculated by the following formula:

[0082]

[0083] wherein, denotes the element value of the i-th row and the j-th column in the Gram matrix G, which reflects the correlation between the time step and the angle corresponding to the time step denotes the angle value after angle encoding of the i-th feature parameter at the time step denotes the angle value after angle encoding of the i-th feature parameter at the time step denotes the angle value after angle encoding of the i-th feature parameter at the time step

[0084] S203, the processing device inputs the preprocessed source domain data into a convolutional neural network-long short-term memory network (CNN-LSTM) hybrid model for pre-training, to obtain a pre-training model.

[0085] The preprocessed source domain data is structured data formed after normalization (eliminating dimensional differences), dimensionality reduction (such as converting the Gram angle field algorithm into two-dimensional image features), etc. of the source domain data (occupant collision data in a standard posture collision scene) for the input of the adaptive model.

[0086] Convolutional Neural Network (CNN) is a kind of deep learning model good at extracting spatial features, which captures local patterns in data through convolution kernel sliding, such as the spatial correlation structure of the collision feature matrix.

[0087] Long Short-Term Memory (LSTM) is a kind of recurrent neural network good at processing time series data, which learns the long-term dependence of sequence data through the gating mechanism, such as the dynamic evolution law of collision acceleration and spatial coordinates.​​​​​​

[0088] The CNN-LSTM hybrid model is a deep learning model that combines CNN and LSTM architectures. It first extracts spatial features of the collision data using CNN, and then captures the temporal dynamic correlation of the features using LSTM, achieving multi-dimensional modeling of the collision scene.

[0089] Pre-training is the initialization and training of the model on source domain data (standard posture scenes with sufficient data and complete labels). The model learns the general feature representation and damage correlation rules in this field.

[0090] The pre-trained model is a model obtained after pre-training, and its parameters have been optimized on the source domain task. It has the ability to predict occupant collision damage levels in standard postures, and is the basis for subsequent transfer learning.

[0091] The processing device inputs the source domain data that has been normalized and dimensionally reduced into the CNN-LSTM hybrid model. It uses the spatial feature extraction capability of CNN and the temporal dependence modeling capability of LSTM to iteratively optimize the model parameters on the source domain data. The model learns the mapping rules between collision features and occupant head injury levels in standard postures, and finally obtains a pre-trained model with stable performance in standard posture scenarios. This process lays the foundation for subsequent fine-tuning of the model using a small amount of high-angle posture data and realization of cross-domain knowledge transfer, solving the problem of model training caused by the scarcity of high-angle posture data.

[0092] The CNN network structure diagram is shown in Figure 2 . The convolution layer uses a convolution kernel to perform convolution operations on the input data and passes the results to the next layer of the network. The expression is:

[0093]

[0094] wherein, is the output value of the th neuron in the th layer network, represents the output value of the th neuron in the th layer network, is the number of neurons connecting the previous layer and the current layer, is the weight parameter of the th neuron in the previous layer and the th neuron in the current layer in the th layer network, is the bias parameter of the th neuron in the th layer network, represents the activation function.

[0095] After the convolutional layers extract the nonlinear features of the data, the Recurrent Neural Network (ReLU) activation function is used to enhance the model's feature representation capability. Simultaneously, to reduce overfitting in CNNs, pooling layers are introduced into the network structure. Pooling layers not only alleviate the overfitting problem but also preserve the main features of the output data. In this application, max pooling layers are used. The expression is:

[0096]

[0097] in, It is the first In a layered network, the first The feature in the first Pooling output values ​​under each pooling region Within the coverage area of ​​the pooling core, the first In the layer network The first feature element, It is the first Pooling results for each pooling region.

[0098] The fully connected layer, serving as the classification module of the convolutional neural network model, uses the softmax function to generate a probability distribution for the extracted nonlinear activation features. The expression is as follows:

[0099]

[0100] in, Indicates the first in the output layer The original output value of each neuron. It is the number of damage levels. After processing by the softmax function, the first The probability output results corresponding to each neuron. Indicates the first in the output layer The original output value of each neuron.

[0101] LSTM network structure as follows Figure 3 As shown, the LSTM network uses input gates, forget gates, and output gates to control the prior state, current input, and current memory, thereby improving the hidden layer of a conventional RNN, and uses activation functions to compute the hidden layer state.

[0102] The input gate controls the information transmitted to the network at each time step, and its calculation formula is as follows:

[0103]

[0104] in, It is the input gate matrix at time step t. is an activation function, is a weight matrix of the input gate, is a bias matrix of the input gate, is the input at time step t, is the hidden state at time step t-1.

[0105] The forget gate determines whether information is retained according to the previous output and the current input, and the calculation formula is:

[0106]

[0107] wherein, is the forget gate matrix at time step t, is a weight matrix of the forget gate, is a bias matrix of the forget gate.

[0108] The output gate outputs information, and the calculation formula is:

[0109]

[0110] wherein, is the output gate matrix at time step t, is a weight matrix of the output gate, is a bias matrix of the output gate.

[0111] The cell state represents the cell state, which is a candidate hidden state generated from the current input and the previous hidden state, and the calculation formula is:

[0112]

[0113]

[0114] wherein, is the cell state at time step t, denotes the cell state at time step t-1, denotes the candidate cell state at time step t, denotes the weight matrix corresponding to the candidate cell state, is a bias parameter corresponding to the candidate cell state, is a hyperbolic tangent activation function.

[0115] The hidden state is the time sequence feature extracted by the LSTM, and the calculation formula is:

[0116]

[0117] wherein, denotes the hidden state at time step t.

[0118] S204, the processing device freezes the parameters of the CNN part in the pre-trained model, fine-tunes the LSTM part and the fully connected layer in the pre-trained model using the pre-processed target domain data, and obtains a transfer learning model of the target posture.

[0119] Specifically, the processing device keeps the weight parameters of the convolutional layer and the pooling layer in the CNN part of the pre-trained model unchanged; inputs the pre-processed target domain data into the pre-trained model to obtain a prediction result; and updates the weight parameters of the LSTM part and the fully connected layer through a back propagation algorithm based on the loss between the prediction result and the real label.

[0120] The processing device fixes the weight parameters of the convolutional layer and the pooling layer in the CNN part of the pre-trained model, because these two layers have learned the general spatial features of the collision scene (such as the position association rule of the restraint system and the occupant) from a large amount of source domain data, and do not need to be retrained, but can be directly reused.

[0121] The pre-processed target domain data (high-angle posture collision feature data) is input into the pre-trained model, and the model extracts spatial features based on the frozen CNN, and then outputs a preliminary prediction result through the unfrozen LSTM part and the fully connected layer.

[0122] Based on the loss value of the prediction result and the real label, the gradient is calculated through the back propagation algorithm, and only the weight parameters of the LSTM part (responsible for capturing the time sequence dynamic difference of the collision features under the high-angle posture) and the fully connected layer (responsible for feature fusion and prediction mapping) are updated, so that the model quickly adapts to the specific scene of the high-angle posture, while reusing the source domain knowledge, and avoiding the overfitting of the model caused by the scarcity of the target domain data.

[0123] Suppose the CNN part in the pre-trained model is , the LSTM part is , the fully connected layer is , and the input features of the target domain are , then the prediction result is:

[0124]

[0125] wherein, represents the frozen weight parameters of the CNN part, represents the initial weight parameters of the LSTM part, represents the initial weight parameters of the fully connected layer, represents the prediction result of the model to the target domain input.

[0126] By back propagation, the loss function is minimized, only and are updated, and the expression is:

[0127]

[0128] wherein, denote the optimal weight parameters of the updated LSTM and fully connected layer, and the updating process satisfies remains unchanged.

[0129] When the injury level prediction accuracy of the transfer learning model on the validation set is greater than or equal to the accuracy threshold, the fine-tuning training is stopped, and the transfer learning model of the target posture is obtained.

[0130] The validation set is an independent data set (not involved in model training) divided from the target domain data, which is used to objectively evaluate the generalization ability of the model in the fine-tuning process and avoid overfitting.

[0131] The injury level prediction accuracy is the matching ratio of the prediction result of the model on the validation set and the true injury level label.

[0132] The accuracy threshold is a preset model performance judgment threshold, for example, the accuracy threshold is 90%, which is determined by engineering requirements or industry standards, and is used to measure whether the model meets the practical accuracy requirements.

[0133] During the fine-tuning process of the target domain data, the processing device inputs the validation set data into the model at regular intervals to calculate the injury level prediction accuracy; the accuracy is compared with the accuracy threshold in real time, and when the accuracy is greater than or equal to the accuracy threshold, it indicates that the model has learned the collision damage law of the large-angle posture sufficiently, and the generalization ability meets the practical requirements, at which time the fine-tuning is stopped to avoid overfitting caused by overtraining; the final obtained model is the transfer learning model adapted to the target posture, which can be directly used for injury level prediction in large-angle scenarios. This mechanism takes into account the model accuracy and training efficiency, ensuring the reliability and engineering practicability of the output model.

[0134] S205, if the riding angle is greater than the angle threshold, input the parameters corresponding to the target posture into the transfer learning model to output the occupant head injury level in the target posture; if the riding angle is less than or equal to the angle threshold, input the parameters corresponding to the standard posture into the pre-trained model to output the occupant head injury level in the standard posture.

[0135] The processing device first acquires the actual riding angle of the current seat and compares it with the angle threshold; if the riding angle is greater than the threshold (determined as a large-angle target posture), the preprocessed parameters corresponding to the posture are input into the transfer learning model, and the model outputs the occupant head injury level in the posture based on the optimized parameters adapted to the large-angle scenario.

[0136] If the riding angle is less than or equal to the angle threshold (determined as a standard posture), the preprocessed parameters corresponding to the posture are input into the pre-trained model, and the model outputs the injury level in the scenario based on the general knowledge of the standard posture.

[0137] The method realizes accurate adaptation of the double model through pose classification, guarantees prediction stability under the standard pose, solves the scene adaptation problem of large-angle pose, and realizes damage level prediction requirements covering all working conditions.

[0138] The method also includes the case of unfreezing the CNN part parameters in the pre-trained model, as follows:

[0139] The processing device monitors the domain difference between the feature distribution of the target domain data and the feature distribution of the source domain data in the fine-tuning process; when the domain difference is greater than the difference threshold, the deep convolutional layer parameters of the CNN part in the pre-trained model are unfrozen; and based on the domain difference, the learning rates of the unfrozen deep convolutional layer parameters, the LSTM part parameters and the fully connected layer parameters are adjusted respectively.

[0140] The domain difference is a quantitative indicator that measures the difference in feature distribution between the source domain (standard pose) data and the target domain (large-angle pose) data. For example, the distance (such as maximum mean difference MMD) between the two domain data distributions can be calculated to measure it. The greater the difference, the more different the collision characteristics of the two scenes, and the more difficult it is to directly reuse the source domain knowledge.

[0141] The difference threshold is a preset critical value for determining whether the domain difference is large enough to require adjustment of the training strategy. When the monitored domain difference exceeds this threshold, the subsequent parameter unfreezing and learning rate adjustment operations are triggered.

[0142] The deep convolutional layer is the convolutional layer close to the output end in the CNN model. These layers usually learn more abstract and higher-level features. Compared with shallow layers that learn basic textures, deep features are more task-specific, so when the target domain and the source domain differ greatly, they need to be adjusted to adapt to the new task.

[0143] The learning rate is a hyperparameter that controls the parameter update amplitude in model training. The larger the learning rate, the faster the parameter update, but it may lead to instability; the smaller the learning rate, the more stable the training, but the convergence speed may slow down.

[0144] In the model fine-tuning stage, the processing device does not use a fixed and unchanged training strategy, but continuously calculates and monitors the distribution difference between the source domain and the target domain data in the model feature space, i.e. the domain difference, to dynamically perceive the adaptation degree of the two domains.

[0145] When the monitored inter-domain difference exceeds the preset difference threshold, it indicates that the universal features learned by the pre-trained model from the source domain and the specificity features of the target domain have a significant gap. At this time, if the initial strategy of only training the LSTM part and the fully connected layer is still followed, the model will have difficulty in fully capturing the key feature patterns of the large-angle collision scenario, thereby leading to insufficient prediction accuracy of the damage level.

[0146] To solve this problem, the processing device triggers an adaptive adjustment mechanism to optimize the adaptation ability of the model to the target domain. Specifically, the adjustment mechanism first unfreezes the CNN deep convolutional layers that were originally frozen in the pre-trained model. These layers are responsible for extracting high-level abstract features of the data, and their parameters can be updated based on the target domain data after being unfrozen, so that the model can learn more feature representations that are more suitable for large-angle collision scenarios, breaking through the adaptation limitations of universal features.

[0147] At the same time, in order to balance the training stability and convergence efficiency, the processing device differentially adjusts the learning rate of different network layers according to the specific value of the inter-domain difference: for the CNN deep convolutional layers that are more sensitive to the inter-domain difference, a smaller learning rate is set to achieve fine parameter optimization and avoid damaging the source domain universal knowledge due to excessive adjustment; for the LSTM part and the fully connected layer, a relatively larger learning rate is set to quickly adapt to the time series features of the target domain and the prediction mapping requirements.

[0148] After the adjustment is completed, the model continues to fine-tune on the target domain data. Through this method of dynamic monitoring, threshold triggering, hierarchical unfreezing, and adaptive parameter adjustment, the model can balance the reuse of mature knowledge from the source domain and the learning of new knowledge from the target domain, effectively overcome the performance bottleneck caused by cross-domain differences, and ultimately achieve better generalization ability and prediction accuracy in the large-angle posture collision damage prediction task.

[0149] Based on the above content description, the present application has the following beneficial effects:

[0150] In view of the problem that large-angle posture collision test data is scarce and traditional prediction models based on standard sitting posture are invalid due to the dramatic change in the interaction between the occupant and the restraint system, the present application uses a transfer learning strategy to fully reuse the damage-related knowledge contained in the rich source domain data under the standard posture, and fine-tunes the model with a small amount of target domain data, greatly reducing the dependence on expensive physical tests or high-fidelity simulation data for large-angle postures. Under the premise of controllable data cost, the prediction accuracy of the occupant head injury level under large-angle postures is improved.

[0151] Further, the pre-training model and the transfer learning model are cooperatively designed in the application, and the corresponding prediction model is adaptively matched according to different seat sitting angles. When the sitting angle does not exceed the threshold, the pre-training model trained by sufficient source domain data is directly used to guarantee the prediction accuracy in the standard posture. When the sitting angle exceeds the threshold, the transfer learning model adapted to the large-angle scene is enabled, thereby realizing full-scene coverage from the regular sitting posture to the large-angle lying posture, and meeting the safety evaluation requirements of the diversified seat configuration of intelligent and electric vehicles.

[0152] Further, the application introduces an inter-domain difference monitoring mechanism. When the feature distribution difference between the target domain and the source domain is large, the CNN deep layer parameters are dynamically unfrozen and the learning rate is adjusted, thereby avoiding model fine-tuning oscillation caused by too large inter-domain difference, fully excavating the learning ability of the model to specific features of large-angle postures, and controlling the fine-tuning termination time through the accuracy threshold, thereby improving the model training efficiency while guaranteeing the prediction accuracy.

[0153] Further, compared with the traditional method relying on a large number of physical tests or complex dynamic modeling, the transfer learning framework based on the CNN-LSTM hybrid model of the application greatly shortens the cycle of safety performance evaluation of large-angle seats, reduces the research and development cost, and provides an efficient and accurate safety evaluation tool for the large-angle seat configuration of high-end passenger vehicles and autonomous vehicles, thereby helping to improve the riding comfort of related vehicle models and build a collision safety line.

[0154] The above is combined with Figure 1 The prediction method for the head injury level of the occupant in a collision provided by the embodiments of the application is described in detail, and the device and the apparatus provided by the embodiments of the application will be introduced below with reference to the accompanying drawings.

[0155] As Figure 4 shown, the figure is a schematic diagram of a prediction device for the head injury level of the occupant in a collision provided by an embodiment of the application, and the device comprises:

[0156] The acquisition module 301 is configured to acquire source domain data and target domain data. The source domain data is occupant collision data in a standard posture, and the target domain data is occupant collision data in a target posture. The sitting angle corresponding to the target posture is greater than an angle threshold, and the sample quantity of the target domain data is less than the sample quantity of the source domain data.

[0157] The data processing module 302 is configured to preprocess the source domain data and the target domain data, input the preprocessed source domain data into a convolutional neural network-long short-term memory network (CNN-LSTM) hybrid model for pre-training, obtain a pre-training model, freeze parameters of a CNN part in the pre-training model, fine-tune an LSTM part and a full connection layer in the pre-training model by using preprocessed target domain data, and obtain a transfer learning model of a target posture.

[0158] The prediction module 303 is configured to input parameters corresponding to the target posture into the transfer learning model if the riding angle is greater than the angle threshold, and output a head injury level of an occupant in the target posture.

[0159] If the riding angle is less than or equal to the angle threshold, the prediction module 303 is configured to input parameters corresponding to the standard posture into the pre-training model, and output a head injury level of an occupant in the standard posture.

[0160] Optionally, the data processing module 302 is configured to perform normalization processing on the source domain data and the target domain data respectively, and perform dimension reduction processing on the normalized source domain data and the normalized target domain data.

[0161] Optionally, the data processing module 302 is configured to perform dimension reduction processing on the normalized source domain data and the normalized target domain data by using a Gram angle field algorithm.

[0162] Optionally, the data processing module 302 is configured to keep weight parameters of convolutional layers and pooling layers in the CNN part of the pre-training model unchanged.

[0163] The data processing module 302 is configured to input the preprocessed target domain data into the pre-training model, and obtain a prediction result.

[0164] Based on a loss between the prediction result and a real label, the data processing module 302 is configured to update weight parameters of the LSTM part and the full connection layer by using a back propagation algorithm.

[0165] Optionally, the data processing module 302 is further configured to stop fine-tuning training when an injury level prediction accuracy of the transfer learning model on a verification set is greater than or equal to an accuracy threshold, and obtain the transfer learning model of the target posture.

[0166] Optionally, the data processing module 302 is further configured to monitor a domain difference degree between a feature distribution of the target domain data and a feature distribution of the source domain data in the fine-tuning process.

[0167] When the domain difference degree is greater than a difference degree threshold, the data processing module 302 is configured to unfreeze deep convolutional layer parameters of the CNN part in the pre-training model.

[0168] Based on the domain difference degree, the data processing module 302 is configured to respectively adjust learning rates of the unfrozen deep convolutional layer parameters, LSTM part parameters and full connection layer parameters.

[0169] The prediction device for occupant head impact injury severity according to the embodiments of the present application can correspond to performing the methods described in the embodiments of the present application, and the above-mentioned other operations and / or functions of each module / unit of the prediction device for occupant head impact injury severity are respectively implemented in order to realize Figure 1 the corresponding flow of each method in the illustrated embodiments, which will not be described here again for brevity.

[0170] The embodiments of the present application also provide a computing device. As Figure 5 shown, the figure is a schematic diagram of a computing device provided by the embodiments of the present application, which includes a bus 701, a processor 702, a communication interface 703 and a memory 704. The processor 702, the memory 704 and the communication interface 703 communicate through the bus 701.

[0171] The bus 701 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0172] The processor 702 can be any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP) processor, etc.

[0173] The communication interface 703 is used for external communication.

[0174] The memory 704 can include a volatile memory (volatile memory), such as a random access memory (RAM). The memory 704 can also include a non-volatile memory (non-volatile memory), such as a read-only memory (ROM), a flash memory, a hard disk drive (HDD) or a solid state drive (SSD).

[0175] The executable code stored in the memory 704 is executed by the processor 702 to perform the aforementioned method for predicting the head injury criterion of the occupant in a collision.

[0176] Specifically, in the case of the embodiment shown, and Figure 4 In the case of the embodiment described, and Figure 4 In the case of the embodiment described, and Figure 4 The software or program code required for the functions of each module / unit in the foregoing method for predicting the head injury criterion of the occupant in a collision can be partially or entirely stored in the memory 704. The processor 702 executes the program code corresponding to each unit stored in the memory 704 to perform the aforementioned method for predicting the head injury criterion of the occupant in a collision.

[0177] The embodiment of the present application also provides a computer readable storage medium. The computer readable storage medium can be any available medium or data storage device that can be used to store data that can be accessed by a computing device. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk), etc. The computer readable storage medium includes instructions that instruct the computing device to perform the aforementioned method for predicting the head injury criterion of the occupant in a collision.

[0178] The embodiment of the present application also provides a computer program product, which includes one or more computer instructions. When the computer instructions are loaded and executed on a computing device, the flow or function described in the embodiment of the present application is generated in whole or in part.

[0179] The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer or data center to another website, computer or data center through wired (such as coaxial cable, optical fiber, digital subscriber line) or wireless (such as infrared, wireless, microwave, etc.) mode.

[0180] The computer program product is executed by a computer, and the computer executes any of the aforementioned methods for predicting the head injury criterion of the occupant in a collision. The computer program product can be a software installation package, and when any of the aforementioned methods for predicting the head injury criterion of the occupant in a collision is needed, the computer program product can be downloaded and executed on the computer.

[0181] The description of the flow or structure corresponding to each of the above figures has its own emphasis, and the part not described in detail in a certain flow or structure can be referred to the related description of other flows or structures.

[0182] The above description is only the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any change or replacement within the technical scope disclosed in the present application should be covered in the protection scope of the present application.

Claims

1. A method of predicting a head injury criterion of an occupant in a collision, characterized by, The method comprises: obtaining source domain data and target domain data, the source domain data being occupant crash data in a standard posture, the target domain data being occupant crash data in a target posture, the target posture corresponding to a ride angle greater than an angle threshold, and the number of samples of the target domain data being less than the number of samples of the source domain data; preprocessing the source domain data and the target domain data; inputting the preprocessed source domain data into a convolutional neural network-long short-term memory network (CNN-LSTM) hybrid model for pretraining to obtain a pretraining model; freezing the parameters of the CNN part of the pretraining model, fine-tuning the LSTM part and the fully connected layer of the pretraining model using the preprocessed target domain data to obtain a transfer learning model for the target posture; The CNN part in the pre-trained model is , the LSTM part is , the full connection layer is , the input feature of the target domain is , and the prediction result is: wherein, denote the initial weight parameters of the CNN part, denote the initial weight parameters of the LSTM part, denote the initial weight parameters of the fully connected layer, denote the prediction results of the model on the target domain input; minimizing the loss function by backpropagation , only update with , the expression is: wherein, denotes the optimal weight parameters of the updated LSTM, fully connected layer after fine-tuning, and the updating process satisfies remains unchanged; monitoring the domain difference between the feature distribution of the target domain data and the feature distribution of the source domain data during fine-tuning; when the domain difference is greater than a difference threshold, unfreezing the deep convolutional layer parameters of the CNN part of the pretraining model; based on the domain difference, adjusting the learning rates of the unfrozen deep convolutional layer parameters, the LSTM part parameters and the fully connected layer parameters, respectively; if the ride angle is greater than the angle threshold, inputting the parameters corresponding to the target posture into the transfer learning model to output the occupant head injury level in the target posture; if the ride angle is less than or equal to the angle threshold, inputting the parameters corresponding to the standard posture into the pretraining model to output the occupant head injury level in the standard posture.

2. The method of claim 1, wherein, The preprocessing of the source domain data and the target domain data comprises: normalizing the source domain data and the target domain data, respectively; dimensionality reduction processing of the normalized source domain data and the target domain data.

3. The method of claim 2, wherein, The dimensionality reduction processing of the normalized source domain data and the target domain data comprises: dimensionality reduction processing of the normalized source domain data and the target domain data by Gram angle field algorithm.

4. The method of claim 1, wherein, The freezing of the parameters of the CNN part of the pretraining model and the fine-tuning of the LSTM part and the fully connected layer of the pretraining model using the preprocessed target domain data comprise: keeping the weight parameters of the convolutional layer and the pooling layer of the CNN part of the pretraining model unchanged; inputting the preprocessed target domain data into the pretraining model to obtain a prediction result; updating the weight parameters of the LSTM part and the fully connected layer by a back propagation algorithm based on the loss between the prediction result and the real label.

5. The method of claim 1, wherein, The method further comprises: when the injury level prediction accuracy of the transfer learning model on the validation set is greater than or equal to an accuracy threshold, stopping fine-tuning training to obtain the transfer learning model for the target posture.

6. The method of claim 1, wherein, The occupant crash data comprises the spatial coordinates of multiple fixed points of a safety belt, the spatial coordinates of a dummy H point, a crash acceleration waveform and a labeled occupant head injury level.

7. A device for predicting the head injury criterion of an occupant in a collision, characterized by The device comprises: an acquisition module configured to obtain source domain data and target domain data, the source domain data being occupant crash data in a standard posture, the target domain data being occupant crash data in a target posture, the target posture corresponding to a ride angle greater than an angle threshold, and the number of samples of the target domain data being less than the number of samples of the source domain data; The data processing module is used to preprocess the source domain data and target domain data; input the preprocessed source domain data into a hybrid Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) model for pre-training to obtain a pre-trained model; freeze the parameters of the CNN part in the pre-trained model; and fine-tune the LSTM part and fully connected layers in the pre-trained model using the preprocessed target domain data to obtain a transfer learning model for the target pose; the CNN part in the pre-trained model is... The LSTM part is The fully connected layer is The input features of the target domain are The prediction result is: wherein, denote the initial weight parameters of the CNN part, denote the initial weight parameters of the LSTM part, denote the initial weight parameters of the fully connected layer, denote the prediction results of the model on the target domain input; minimizing the loss function by backpropagation , only update with , the expression is: wherein, denotes the optimal weight parameters of the LSTM, fully connected layer after fine-tuning, and the update process satisfies remains unchanged; monitor an inter-domain difference between a feature distribution of the target domain data and a feature distribution of the source domain data in the fine-tuning process; when the inter-domain difference is greater than a difference threshold, unfreeze deep convolutional layer parameters of a CNN part of the pre-trained model; based on the inter-domain difference, adjust learning rates of the unfrozen deep convolutional layer parameters, LSTM part parameters and full connection layer parameters respectively; The prediction module is configured to: if the riding angle is greater than an angle threshold, input parameters corresponding to the target posture into the transfer learning model to output a head injury level of the occupant in the target posture; and if the riding angle is less than or equal to the angle threshold, input parameters corresponding to the standard posture into the pre-trained model to output a head injury level of the occupant in the standard posture.

8. A computing device, comprising: comprise a memory and a processor; The memory stores one or more computer programs comprising instructions, which, when executed by the processor, cause the computing device to perform the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is configured to store a computer program for performing the method of any one of claims 1 to 6.

Images