Brain injury prediction method and prediction system based on multivariate feature fusion framework

Abstract

The invention provides a brain injury prediction method and prediction system based on a multivariate feature fusion framework, and relates to the technical field of deep learning. In order to solve the problem that dynamic characteristics and static characteristics cannot be effectively fused when multivariate data damage prediction is carried out by adopting a nonlinear regression method in the prior art, dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static influence effect parameters and brain damage degree data under various working conditions of a vehicle are obtained. A dynamic mechanical effect parameter is used as output to train a dynamic prediction model, accurate extraction of a dynamic feature time sequence rule can be ensured, a static influence effect parameter is used as output to ensure extraction of a static feature structured rule, and finally a dynamic feature vector and a static feature vector are used as input of a fusion prediction model. Fusion of a dynamic time sequence rule and a static structured rule is achieved, it is ensured that the two types of features are fully fused, and the prediction result is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a brain injury prediction method and prediction system based on a multi-feature fusion framework. Background Art

[0002] During a vehicle collision, the acceleration generated directly determines the intensity of the occupant's head movement, resulting in varying degrees of brain damage under different conditions. The impact of a vehicle collision on brain damage involves multiple factors, which affect the extent of brain damage by acting directly or indirectly on the occupant's head. Predicting brain damage can not only provide guidance for immediate post-collision rescue efforts but also drive vehicle safety design optimization, assist in accident decision-making, and provide early warning. When predicting brain damage caused by vehicle collisions, the influence of multiple characteristics must be considered.

[0003] Currently, methods for predicting damage based on multivariate features generally employ machine learning algorithms. For example, nonlinear regression is used for multivariate damage prediction. Nonlinear regression involves constructing a nonlinear regression model to explore the relationship between multiple independent variables and a dependent variable, thereby predicting damage. However, while nonlinear regression can fit the nonlinear relationship between features and damage, due to the significant differences in the temporal nature of dynamic features and the structured nature of static features, it struggles to account for the core laws of both types of features. It also tends to overlook the synergistic effects of multivariate features and fails to effectively integrate dynamic and static features.

[0004] Therefore, developing a brain injury prediction method based on a multivariate feature fusion framework is of great significance for improving the accuracy of brain injury prediction. Summary of the Invention

[0005] To address the problem that the existing technology of using nonlinear regression method to predict multivariate data damage cannot effectively integrate dynamic features and static features, the present invention proposes a brain injury prediction method based on a multivariate feature fusion framework, which specifically includes the following steps: Obtain the vehicle's dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions; Training a dynamic prediction model using the dynamic feature data as input and the dynamic mechanical effect parameters as output; Taking the static feature data as input and the static impact effect parameters as output, training a static prediction model; The dynamic prediction model outputs a dynamic feature vector according to the dynamic feature data, and the static prediction model outputs a static feature vector according to the static feature data; Taking the dynamic feature vector and the static feature vector as input and the brain injury degree data as output, training to obtain a fusion prediction model; The dynamic feature data and / or static feature data acquired in real time are used as input to the fusion prediction model to obtain a brain injury prediction result.

[0006] Furthermore, the dynamic feature data is used as input and the dynamic mechanical effect parameters are used as output to train the dynamic prediction model, including: filtering and standardizing the acquired dynamic feature data; and the processed dynamic feature data is used as input and the dynamic mechanical effect parameters are used as output to train the dynamic prediction model.

[0007] Furthermore, the filtering processing of the acquired dynamic feature data includes: decomposing the dynamic feature data into wavelet decomposition coefficients corresponding to the frequency according to the frequency based on the wavelet decomposition method; comparing the wavelet decomposition coefficients with a set threshold; eliminating the wavelet decomposition coefficients less than the set threshold, and setting the wavelet decomposition coefficients greater than or equal to the set threshold as effective wavelet coefficients; and generating the dynamic feature data after filtering optimization according to the effective wavelet coefficients.

[0008] Furthermore, the standardization processing includes: calculating a mean and a standard deviation based on the dynamic feature data after the filtering optimization; and performing standardization processing on the dynamic feature data after the filtering optimization according to the mean and the standard deviation to obtain standardized dynamic feature data.

[0009] Furthermore, the training dynamic prediction model includes: inputting the standardized dynamic feature data into the time series convolution model; the time series convolution model outputs a dynamic feature response value based on the input dynamic feature data; and adjusting the parameters of the time series convolution model based on the difference between the dynamic feature response value and the dynamic mechanical effect parameter to determine the dynamic prediction model.

[0010] Furthermore, the static feature data is used as input and the static influence effect parameters are used as output to train a static prediction model, including: normalizing the acquired static feature data; and using the normalized static feature data as input and the static influence effect parameters as output to train a static prediction model.

[0011] Furthermore, the training of the static prediction model includes: inputting the normalized static feature data into the gradient boosting tree model; the gradient boosting tree model outputs a static feature response value based on the input static feature data; and adjusting the parameters of the gradient boosting tree model based on the difference between the static feature response value and the static influence effect parameter to determine the static prediction model.

[0012] Furthermore, after obtaining the brain injury prediction result, it also includes: calculating the contribution of the static feature data to the prediction result, specifically including: based on a decision structure tree model, determining the parent node and child node of the static feature data, as well as the predicted value corresponding to the parent node and the predicted value corresponding to the child node; inputting the parent node, the child node, the predicted value corresponding to the parent node and the predicted value corresponding to the child node into a contribution calculation model to obtain the contribution.

[0013] Furthermore, the dynamic feature data and / or static feature data acquired in real time are used as input and input into the fusion prediction model to obtain a brain injury prediction result, specifically including: collecting the dynamic feature data and / or static feature data of the vehicle in real time through sensors; filtering and normalizing the dynamic feature data; normalizing the static feature data; inputting the processed dynamic feature data and / or static feature data into the fusion prediction model to obtain a brain injury prediction result; and issuing an alarm when the brain injury prediction result exceeds a set threshold.

[0014] The present invention also provides a brain injury prediction system based on a multi-feature fusion framework. The prediction system adopts the brain injury prediction method based on a multi-feature fusion framework as described in any of the above items, and specifically includes the following modules: A parameter acquisition module is used to obtain the vehicle's dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions; a dynamic prediction model training module, connected to the parameter acquisition module, for taking the dynamic feature data as input and the dynamic mechanical effect parameters as output to train a dynamic prediction model, wherein the dynamic prediction model outputs a dynamic feature vector according to the dynamic feature data; a static prediction model training module, connected to the parameter acquisition module, for taking the static feature data as input and the static impact effect parameters as output to train a static prediction model, wherein the static prediction model outputs a static feature vector according to the static feature data; A fusion prediction model training module is connected to the dynamic prediction model training module and the static prediction model training module, and is used to take the dynamic feature vector and the static feature vector as input and the brain injury degree data as output to train a fusion prediction model, wherein the fusion prediction model predicts brain injury based on the dynamic feature data and / or static feature data acquired in real time.

[0015] Compared with the prior art, the present invention has the following beneficial effects: First, by acquiring the vehicle's dynamic feature data, static feature data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions, the dynamic feature data is used as input and the dynamic mechanical effect parameters are used as output to train the dynamic prediction model; the static feature data is used as input and the static impact effect parameters are used as output to train the static prediction model, and the dynamic feature vector output by the dynamic prediction model and the static feature vector output by the static prediction model are used as input and the brain injury degree data are used as output to train a fusion prediction model, and the prediction results are obtained based on the fusion prediction model. Training the dynamic prediction model with the dynamic mechanical effect parameters as output can ensure the accurate extraction of the temporal laws of the dynamic features, and using the static impact effect parameters as output can ensure the extraction of the structured laws of the static features. Finally, the dynamic feature vectors and the static feature vectors are used as input to the fusion prediction model to achieve the fusion of dynamic temporal laws and static structured laws, ensuring that the two types of features are fully integrated, making the prediction results more accurate; Second, the acquired dynamic feature data is filtered and standardized. Because dynamic feature data is easily interfered with by objective factors such as vehicle vibration and structural noise during the acquisition process, if directly used for fusion, the model training results will learn the patterns of noise, affecting the prediction results. Through filtering processing based on wavelet decomposition, error interference such as vibration can be accurately eliminated, retaining the core information directly related to brain injury in the dynamic feature data, avoiding the impact of interfering data on model training, and ensuring that the fusion of dynamic and static feature data is based on real and valid data, further improving the reliability of the fused prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 is a flowchart of a brain injury prediction method based on a multi-feature fusion framework provided by an embodiment of the present invention; Figure 2 is a calculation flow chart of wavelet decomposition provided by an embodiment of the present invention; Figure 3 is a relationship diagram between contribution and prediction results provided by an embodiment of the present invention; Figure 4 The figure is a schematic structural diagram of a prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0019] The specific embodiments of the present invention are described below.

[0020] To address the problem that existing technologies using nonlinear regression for multivariate damage prediction cannot effectively integrate dynamic and static features, the present invention obtains dynamic and static feature data of the vehicle, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions. Using the dynamic mechanical effect parameters as output, a dynamic prediction model is trained to ensure the accurate extraction of dynamic feature temporal patterns. Using the static impact effect parameters as output, the structured patterns of static features are extracted. Ultimately, the dynamic and static feature vectors are used as inputs for a fusion prediction model to achieve the fusion of dynamic temporal patterns and static structured patterns, ensuring that both types of features are fully integrated and resulting in more accurate prediction results.

[0021] Example 1 The embodiment of the present invention provides a brain injury prediction method based on a multi-feature fusion framework. Figure 1 is a flowchart of a brain injury prediction method based on a multi-feature fusion framework provided by an embodiment of the present invention, such as Figure 1 As shown in FIG, the brain injury prediction method based on the multi-feature fusion framework specifically includes the following steps: S1. Obtain the vehicle's dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions.

[0022] Dynamic feature data refers to mechanical data that changes over time during a vehicle collision. It is time-series data. For example, dynamic feature data includes acceleration and angular velocity. Static feature data is objective and relatively fixed attribute data during a collision. It does not change dynamically over time. For example, static feature data includes the occupant's age, gender, height, and weight, as well as the vehicle's stiffness level, collision location, and collision angle.

[0023] Dynamic mechanical effect parameters refer to the core parameters obtained after processing and extracting dynamic feature data, which are used to characterize the collision energy distribution and dynamic mechanical changes, such as the energy peak position, energy transfer efficiency, acceleration / angular velocity time series curve characteristics, etc. Static impact effect parameters refer to the parameters obtained after training and fusing static feature data, which are used to reflect the influence of static feature data on brain injury, such as the quantitative value of injury tolerance at different ages, the influence value of different collision positions on head force, etc. Brain injury degree data under various working conditions refers to information related to the degree of brain injury of occupants in different collision scenarios. For example, it can be expressed by quantitative values ​​such as the impact force received by the occupants.

[0024] S2. Take the dynamic feature data as input and the dynamic mechanical effect parameters as output to train the dynamic prediction model.

[0025] Specifically, training the dynamic prediction model includes: filtering and standardizing the acquired dynamic feature data; using the processed dynamic feature data as input and the dynamic mechanical effect parameters as output to train the dynamic prediction model.

[0026] The filtering process for the acquired dynamic feature data includes: decomposing the dynamic feature data into wavelet decomposition coefficients corresponding to the frequencies based on the wavelet decomposition method; comparing the wavelet decomposition coefficients with a set threshold; eliminating wavelet decomposition coefficients less than the set threshold, and setting the wavelet decomposition coefficients greater than or equal to the set threshold as valid wavelet coefficients; and generating filtered and optimized dynamic feature data based on the valid wavelet coefficients. The calculation formula is as follows: ; Where, Represents the set threshold, Represents the wavelet decomposition coefficient, each wavelet decomposition coefficient corresponds to a signal in a specific frequency range, is the effective wavelet coefficient, sign(w j,k ) is a sign function used to retain the positive and negative signs of the original coefficients. When dynamic characteristic data containing interference factors such as vibration is input, it is decomposed through the wavelet basis to obtain a series of wavelet decomposition coefficients. When the original wavelet decomposition coefficients The absolute value is greater than or equal to the set threshold When , the core coefficients are calculated by the formula, that is, the effective wavelet coefficients after eliminating noise interference. When the absolute value of the original wavelet decomposition coefficient is less than the threshold When , it is determined to be a noise correlation coefficient and is directly set to 0 to filter out irrelevant interference signals in the dynamic feature data.

[0027] For example, after filtering a large amount of data with different test forms and different test characteristics, multiple groups of empirically filtered data can be obtained. Figure 2 is a calculation flow chart of wavelet decomposition provided by an embodiment of the present invention, such as Figure 2 As shown in the figure, the dynamic feature data is subjected to wavelet decomposition, which includes the selection of wavelet basis, the number of wavelet decomposition layers, the selection of wavelet threshold, the selection of wavelet threshold function, and finally the data after filtering optimization is obtained through wavelet reconstruction. By selecting different wavelet bases (B1, B2, ...) and decomposition layers (N=1, N=2, ...), the initial dynamic feature data T jk Perform wavelet decomposition to obtain wavelet coefficients. Based on the selected wavelet threshold (λ1,λ2,…) and threshold function (Fun(1),Fun(2),…), filter the wavelet coefficients to screen out valid data and calculate the process.

[0028] The standardization process includes: calculating a mean value and a standard deviation based on the dynamic feature data after filtering optimization; and performing standardization process on the dynamic feature data after filtering optimization based on the mean value and the standard deviation to obtain standardized dynamic feature data.

[0029] The mean value calculated from the dynamic feature data after filtering optimization can reflect the central tendency of the feature data. The standard deviation calculated from the dynamic feature data after filtering optimization can reflect the degree of dispersion of the data. The dynamic feature data after filtering optimization is processed by the standardization formula to eliminate the dimensional differences, which is convenient for direct use in dynamic prediction model training. The standardization method uses the following calculation formula: ; For example, if we input a vehicle collision speed sample [32, 45.6, 50.5, 63.5, 81.3] and perform normalization on it, we can get the data shown in Table 1: Table 1 Standardization results

[0030] Training the dynamic prediction model specifically includes: inputting the standardized dynamic feature data into the time series convolution model; the time series convolution model outputs the dynamic feature response value based on the input dynamic feature data; and adjusting the parameters of the time series convolution model based on the difference between the dynamic feature response value and the dynamic mechanical effect parameter to determine the dynamic prediction model. The time series convolution model is expressed as: ; Where X represents the input normalized dynamic feature data, W0 represents the weight matrix of the embedding layer, which is used to perform a linear transformation on the input data X and map it to a high-dimensional feature space. Its dimension is determined by the input data and the output dimension of the embedding layer. b0 represents the bias term of the embedding layer, which is used to adjust the output value of the embedding layer after linear transformation to increase the flexibility of the model. σ represents the activation function, which is used to introduce nonlinearity and enhance the expressive power of the model. H0 represents the output of the embedding layer, that is, the feature representation obtained after processing by the embedding layer, which serves as the input of the subsequent temporal convolution layer. Z i Represents the input of the i-th layer of temporal convolution, W i represents the weight matrix of the i-th layer of temporal convolution, Represents the dilated convolution operation, d represents the dilation rate, which is usually expanded in the form of an exponential of 2. Indicates the The output of the temporal convolution layer. b i Represents the bias term of the i-th layer temporal convolution, which is used to adjust the output value after the convolution operation. s Represents the weight matrix of the residual connection, which is used to Output of the layer Perform linear transformation to achieve residual connection and alleviate the gradient disappearance problem. i Represents the output of the i-th layer of temporal convolution, and the feature representation obtained after residual connection and activation function processing. N represents the total number of layers of temporal convolution, which determines the depth of the model's extraction of temporal features. out Represents the weight matrix of the output layer, which is used to adjust the output H of the last layer of temporal convolution N Perform linear transformation and map to the output dimension. b out Represents the bias term of the output layer, which is used to adjust the output value after the linear transformation of the output layer. out Represents the activation function of the output layer, which is used to map the output values ​​to a specific range. Y represents the final output of the model.

[0031] The temporal convolution model converts data into a processable feature format through the embedding layer. The temporal convolution model processes the input data through causal dilation convolution and residual connection to generate dynamic feature response values. The dynamic feature response value at this time may deviate from the target due to the unoptimized model parameters. The dynamic feature response value is compared with the dynamic mechanical effect parameter and the difference between the two is calculated. Based on the error results, the model parameters, such as the convolution weight matrix W, are adjusted through the backpropagation algorithm. i , bias term b i, allowing the model to gradually learn the relationship between dynamic features and target parameters. This iterative process of input, output response, and error adjustment is repeated until the error between the dynamic feature response value and the dynamic mechanical effect parameter stabilizes within a preset threshold, for example, less than 1%. At this point, the model has mastered the relationship between dynamic features and dynamic mechanical effects, and the final dynamic prediction model is determined. This model can stably output dynamic feature vectors representing the collision energy distribution, which are then used for fusion with static feature vectors.

[0032] S3. Take the static feature data as input and the static impact effect parameters as output to train the static prediction model.

[0033] Specifically, training the static prediction model specifically includes: normalizing the acquired static feature data; using the normalized static feature data as input and the static influence effect parameters as output to train the static prediction model.

[0034] Normalization is the process of converting data to a specific range, eliminating the dimensional differences and numerical range differences between different features, which facilitates subsequent model training and analysis. For example, the acquired static feature data is normalized so that the normalized value eventually falls between [0, 1]. The specific calculation formula is as follows: ; Input a sample of different age distributions [3, 6, 10, 30], and the results obtained after normalization are shown in Table 2: Table 2 Normalization results

[0035] Training the static prediction model specifically includes: inputting the normalized static feature data into the gradient boosting tree model; the gradient boosting tree model outputs a static feature response value based on the input static feature data; and adjusting the parameters of the gradient boosting tree model based on the difference between the static feature response value and the static influence effect parameter to determine the static prediction model. The gradient boosting tree model is expressed as: ; Among them, F M (x) represents the output of the gradient boosting tree model after M iterations, and x represents the input normalized static feature data. F0(x) represents the initial prediction function, which is usually a constant and serves as the starting point for model iteration. M represents the number of iterations of the gradient boosting tree model, that is, the number of decision trees. η represents the learning rate, which is used to control the contribution of each decision tree to the final prediction result, and its value range is usually between 0 and 1. m (x) represents the prediction function of the mth decision tree, which is the basic component of the model. Each decision tree is responsible for learning the previous The residual of each tree. m represents the index of the decision tree, from 1 to M.

[0036] The gradient boosting tree model processes the input static feature data and outputs a static feature response value through a combination of predictions from multiple decision trees. This static feature response value is then compared with the actual static impact effect parameter, and the difference between the two is calculated. Based on the calculated error, the gradient boosting method is used to adjust the parameters of the gradient boosting tree model, such as the number of decision trees, depth, and learning rate, to reduce the prediction error. The above steps are repeated, and the model parameters are continuously optimized until the model's prediction error meets the preset requirements. At this point, the model determined is the static prediction model.

[0037] S4. The dynamic prediction model outputs a dynamic feature vector based on the dynamic feature data, and the static prediction model outputs a static feature vector based on the static feature data.

[0038] The dynamic feature vector is the result of processing the dynamic feature data by the dynamic prediction model. It is the core parameter used to characterize the collision energy distribution and dynamic mechanical changes after filtering optimization and extraction using a time-series convolutional network. Examples include the peak value and duration of the collision acceleration, and the characteristics of the curve of seatbelt force variation over collision time. The static feature vector is the result of processing the static feature data by the static prediction model. It is the core parameter used to characterize the impact of the inherent attributes of the vehicle and occupant on the collision outcome after normalization and extraction using a gradient boosting tree model. Examples include the injury risk level predicted based on vehicle mass and collision angle, and the injury sensitivity predicted based on occupant age and weight.

[0039] S5. Take the dynamic feature vector and the static feature vector as input and the brain injury degree data as output, and train to obtain a fusion prediction model.

[0040] The dynamic feature vector and static feature vector obtained above are used as new input data, and the brain injury degree data is used as the corresponding output label, and are input into the fusion prediction model for training. The fusion prediction model includes a prediction formula and a training formula. The prediction formula is: ; in, It is the prediction method of the fusion prediction model. The specific prediction method is as follows: ; The training formula is: ; Among them, p(x n) represents the set of feature vectors obtained after the nth sample is processed by the dynamic prediction model and the static prediction model, w represents the weight vector of the fusion prediction model, which is used to linearly transform the input features, b is the bias term of the fusion prediction model, which is used to correct the baseline of the prediction result, and σ represents the Sigmoid activation function, which maps the linear transformation result to the interval [0,1]. Represents the final prediction result of the fusion prediction model for the nth sample. Indicates the final prediction method of the fusion prediction model obtained after training, argmin Mmeta The function is to find the prediction method of the fusion prediction model when the subsequent objective function obtains the minimum value Parameter or structure, y n Indicates the true label of the nth sample, which is the degree of brain damage data, x n Represents the input data of the nth sample, which is the original data used to predict the degree of brain damage, including dynamic feature data and static feature data. meta (p(x n )) represents the prediction result of the fusion prediction model for the nth sample, that is, the brain injury degree value predicted based on the dynamic feature vector and the static feature vector. is the objective function, which means that for I samples, the fusion prediction model is calculated The prediction result M for each sample meta (p(x n )) and the actual brain damage degree data y of the sample n The loss L between them is then summed over all samples.

[0041] At the beginning of training, the parameters w and b of the fusion prediction model are randomly initialized, and the prediction formula receives the dynamic feature vector and the static feature vector as input, which is denoted as p(x n ), linearly transform the input features to obtain w T p(x n )+b, the linear transformation result is mapped to the probability interval through the Sigmoid function, and we get In the training formula, the predicted probability is calculated The actual degree of brain damage nThe loss is calculated by summing the losses of all samples to obtain the total loss. The gradient of the total loss with respect to the parameters w and b is calculated using the chain rule. Based on the direction of the gradient, the parameters w and b are adjusted to minimize the loss. Through multiple iterations of optimization, when the error between the model's predictions and the actual brain injury data reaches a preset acceptable range, a trained fusion prediction model is obtained. This model integrates both dynamic and static information, enabling a more comprehensive assessment of the risk of brain injury in vehicle collisions. Compared to models that rely solely on a single eigenvector, it has higher prediction accuracy and reliability.

[0042] S6. Input the dynamic feature data and / or static feature data acquired in real time into the fusion prediction model to obtain a brain injury prediction result. Specifically, the following steps are performed: The vehicle's dynamic feature data and / or static feature data are collected in real time through sensors; the dynamic feature data are filtered and standardized; the static feature data are normalized; the processed dynamic feature data and / or static feature data are input into a fusion prediction model to obtain a brain injury prediction result; when the prediction result exceeds a set threshold, an alarm is issued.

[0043] Onboard sensors collect real-time, time-series mechanical data during a vehicle collision. For example, acceleration data is collected through accelerometers. This data directly reflects the dynamic mechanical changes during a collision. Fixed attribute data, such as occupant age and gender, is preset through the onboard system. Seat sensors can capture height and weight data, and body sensors or the navigation system can capture information such as the vehicle's stiffness rating, impact location, and impact angle. This data remains constant over time during a collision scenario and serves as an inherent attribute for assessing injury risk. Dynamic feature data is filtered and normalized, while static feature data is normalized. The processed dynamic feature data is then input into a dynamic prediction model to generate a dynamic feature vector. The processed static data is then input into a static prediction model to generate a static feature vector. At least one of the dynamic and static feature vectors is then input into the prediction formula of the fused prediction model to generate the final prediction result. If the prediction result exceeds a set threshold, the system automatically triggers an alarm, for example, through in-vehicle audio and visual notifications, seat vibration, or even initiating a call to the back-end rescue system. If the threshold is within the threshold, monitoring continues. Through a comprehensive design encompassing multi-dimensional data collection, data preprocessing, fusion model prediction, and threshold warning, this system achieves real-time, accurate prediction and early warning of brain injuries caused by vehicle collisions. Its core advantage lies in integrating multiple features, eliminating interference, and highlighting key factors. This approach not only addresses the limitations of traditional methods, which rely on single-data sets, low accuracy, and poor timeliness, but also directly contributes to occupant safety warnings and rescue assistance.

[0044] After obtaining the brain injury prediction result, it also includes: calculating the contribution of static feature data to the prediction result, specifically including: based on the decision tree model, determining the parent node, child node, and the prediction value corresponding to the parent node and the prediction value corresponding to the child node of the static feature data; inputting the parent node, child node, the prediction value corresponding to the parent node and the prediction value corresponding to the child node into the contribution calculation model to obtain the contribution. Among them, the decision tree model is a single decision tree in the gradient boosting tree model. Contribution The specific calculation formula is as follows: ; Among them, v j and v parent (j) corresponds to the predicted value of the child node and the parent node in the decision tree, J represents the total number of split nodes in the decision tree, and j represents the sequence number of the split node in the decision tree. For example, for a static feature TZ, covering 1000 data, this data is divided into two child nodes: Node 1 (j=1), covering 600 samples, predicting the degree of damage Node 2 (j=2), covering 400 samples, predicting the degree of damage ; The predicted value of the parent node is The contribution of this feature is for: ; The calculation result is 0.04, which shows the contribution of the corresponding feature to the prediction. Then, the importance of the features can be ranked according to the contribution. When predicting new data samples, low-impact factors can be excluded, which facilitates subsequent design and optimization.

[0045] At the same time, the contribution can also intuitively show the relationship between features and prediction results. Figure 3 is a relationship diagram between contribution and prediction results provided by an embodiment of the present invention, such as Figure 3As shown, the horizontal axis represents the size of the eigenvalue, and the eigenvalue values ​​from left to right gradually increase. The color bar on the right represents the degree of damage, where red represents the greatest degree of damage, blue represents the least degree of damage, and purple represents the degree of damage between the degrees of damage represented by red and blue. For example, multiple samples are collected for each eigenvalue. For the age feature: when the age is young, the damage will be greater. As the age increases, the degree of damage decreases. When the age is too old and the physical condition declines, the degree of damage will also increase accordingly; for the dangerous structure distance feature: when the dangerous structure distance feature value is small, the degree of damage is greater, and when the dangerous structure distance feature value is large, the degree of damage is smaller; for the interior sharpness feature: when the interior sharpness feature value is large, it means that the sharpness is high and the degree of damage is greater. When the interior sharpness feature value is small, it means that the sharpness is low and the degree of damage is smaller.

[0046] Example 2 An embodiment of the present invention further provides a prediction system, which adopts the brain injury prediction method based on the multi-feature fusion framework described in any of the above embodiments. Figure 4 This is a schematic diagram of the structure of a prediction system provided by an embodiment of the present invention, such as Figure 4 As shown, the prediction system includes the following modules: Parameter acquisition module 110, for acquiring dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions of the vehicle; A dynamic prediction model training module 120 is connected to the parameter acquisition module and is used to take the dynamic feature data as input and the dynamic mechanical effect parameters as output to train the dynamic prediction model, wherein the dynamic prediction model outputs a dynamic feature vector based on the dynamic feature data; A static prediction model training module 130 is connected to the parameter acquisition module and is used to take the static feature data as input and the static impact effect parameter as output to train the static prediction model, wherein the static prediction model outputs a static feature vector based on the static feature data; The fusion prediction model training module 140 is connected to the dynamic prediction model training module and the static prediction model training module, and is used to take the dynamic feature vector and the static feature vector as input and the brain injury degree data as output to train a fusion prediction model, wherein the fusion prediction model predicts brain injury based on the dynamic feature data and / or static feature data obtained in real time.

[0047] This embodiment acquires a large amount of multivariate data, eliminates the influence of irrelevant factors, and highlights strongly correlated predictive factors to achieve relatively accurate brain injury prediction. The dynamic feature model architecture is specifically designed for processing time series data, capturing long-term dependencies and predicting time series-related features. The static learning model architecture, on the other hand, incorporates a large number of static features through iterative training, performing regression and classification, and then integrating them with dynamic features. Through a model stacking training method, the two models are synergistically used to predict brain injury. During the fusion architecture learning process, the dynamic prediction model is trained using dynamic mechanical effect parameters as output, ensuring the accurate extraction of temporal patterns of dynamic features. Static impact effect parameters are used as output to ensure the extraction of structural patterns of static features. Ultimately, the dynamic and static feature vectors are used as inputs to the fusion prediction model, achieving the fusion of dynamic temporal patterns and static structural patterns. This fusion architecture-based brain injury prediction method significantly covers the complex factors of brain injury during vehicle collisions, while simultaneously reducing computational cost and time, and improving prediction efficiency and accuracy.

[0048] Applying the prediction system to vehicles enables real-time predictions. Real-time dynamic feature data and actual static feature data are fed into a fusion architecture, outputting real-time risk predictions for brain injury and providing early warnings to passengers, ensuring safer driving and protecting passengers' lives and property. It can also assist with rescue and medical care related to accidents.

[0049] Example 3 An electronic device, comprising: processor and memory; The processor is configured to execute the steps of the brain injury prediction method based on a multi-feature fusion framework as described in any one of Example 1 by calling the program or instructions stored in the memory.

[0050] Example 4 A computer-readable storage medium comprising computer program instructions, wherein the computer program instructions enable a computer to execute the steps of the brain injury prediction method based on a multivariate feature fusion framework as described in any one of Example 1.

[0051] The computer-readable storage medium may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A brain injury prediction method based on a multivariate feature fusion framework, characterized by: include: Obtain the vehicle's dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions; Training a dynamic prediction model using the dynamic feature data as input and the dynamic mechanical effect parameters as output; Taking the static feature data as input and the static impact effect parameters as output, training a static prediction model; The dynamic prediction model outputs a dynamic feature vector according to the dynamic feature data, and the static prediction model outputs a static feature vector according to the static feature data; Taking the dynamic feature vector and the static feature vector as input and the brain injury degree data as output, training to obtain a fusion prediction model; The dynamic feature data and / or static feature data acquired in real time are used as input to the fusion prediction model to obtain a brain injury prediction result.

2. The brain injury prediction method based on the multivariate feature fusion framework according to claim 1 is characterized in that: The method of training a dynamic prediction model by using the dynamic feature data as input and the dynamic mechanical effect parameters as output includes: Filtering and standardizing the acquired dynamic feature data; The processed dynamic feature data is used as input, and the dynamic mechanical effect parameters are used as output to train a dynamic prediction model.

3. The brain injury prediction method based on the multivariate feature fusion framework according to claim 2 is characterized in that: The filtering process of the acquired dynamic feature data includes: Decomposing the dynamic feature data into wavelet decomposition coefficients corresponding to the frequencies according to the frequency based on the wavelet decomposition method; comparing the wavelet decomposition coefficient with a set threshold; Eliminate the wavelet decomposition coefficients that are less than the set threshold, and set the wavelet decomposition coefficients that are greater than or equal to the set threshold as valid wavelet coefficients; Dynamic feature data after filtering optimization is generated according to the effective wavelet coefficients.

4. The brain injury prediction method based on the multivariate feature fusion framework according to claim 3 is characterized in that: The standardization process includes: Calculating the mean and standard deviation based on the dynamic feature data after filtering optimization; The dynamic feature data after filtering and optimization is standardized according to the mean and standard deviation to obtain standardized dynamic feature data.

5. The brain injury prediction method based on the multivariate feature fusion framework according to claim 4 is characterized in that: The training dynamic prediction model includes: Input the normalized dynamic feature data into the temporal convolutional model; The temporal convolution model outputs a dynamic feature response value according to the input dynamic feature data; According to the difference between the dynamic characteristic response value and the dynamic mechanical effect parameter, the parameters of the time series convolution model are adjusted to determine the dynamic prediction model.

6. The brain injury prediction method based on a multivariate feature fusion framework according to claim 1, characterized in that: Taking the static feature data as input and the static impact effect parameters as output, training a static prediction model includes: Normalize the acquired static feature data; The normalized static feature data is used as input, and the static impact effect parameters are used as output to train a static prediction model.

7. The brain injury prediction method based on the multivariate feature fusion framework according to claim 6 is characterized in that: The training of the static prediction model includes: Input the normalized static feature data into the gradient boosting tree model; The gradient boosting tree model outputs a static feature response value according to the input static feature data; According to the difference between the static feature response value and the static influence effect parameter, the parameters of the gradient boosting tree model are adjusted to determine the static prediction model.

8. The brain injury prediction method based on a multi-feature fusion framework according to claim 1 is characterized in that: After obtaining the brain injury prediction result, the method further includes calculating the contribution of the static feature data to the prediction result, specifically including: Based on the decision tree model, determining the parent node and child node of the static feature data, as well as the predicted value corresponding to the parent node and the predicted value corresponding to the child node; The parent node, the child node, the predicted value corresponding to the parent node, and the predicted value corresponding to the child node are input into a contribution calculation model to obtain the contribution.

9. The brain injury prediction method based on a multi-feature fusion framework according to claim 1, characterized in that: The dynamic feature data and / or static feature data acquired in real time are used as input to the fusion prediction model to obtain a brain injury prediction result, which specifically includes: Collecting dynamic characteristic data and / or static characteristic data of the vehicle in real time through sensors; Performing filtering and standardization processing on the dynamic feature data; performing normalization processing on the static feature data; Inputting the processed dynamic feature data and / or static feature data into the fusion prediction model to obtain a brain injury prediction result; When the brain injury prediction result exceeds the set threshold, an alarm is issued.

10. A brain injury prediction system based on a multi-feature fusion framework, characterized by: The prediction system adopts the brain injury prediction method based on the multi-feature fusion framework according to any one of claims 1 to 9, and specifically includes the following modules: A parameter acquisition module is used to obtain the vehicle's dynamic characteristic data, static characteristic data, dynamic mechanical effect parameters, static impact effect parameters, and brain injury degree data under various working conditions; a dynamic prediction model training module, connected to the parameter acquisition module, for taking the dynamic feature data as input and the dynamic mechanical effect parameters as output to train a dynamic prediction model, wherein the dynamic prediction model outputs a dynamic feature vector according to the dynamic feature data; a static prediction model training module, connected to the parameter acquisition module, for taking the static feature data as input and the static impact effect parameters as output to train a static prediction model, wherein the static prediction model outputs a static feature vector according to the static feature data; A fusion prediction model training module is connected to the dynamic prediction model training module and the static prediction model training module, and is used to take the dynamic feature vector and the static feature vector as input and the brain injury degree data as output to train a fusion prediction model, wherein the fusion prediction model predicts brain injury based on the dynamic feature data and / or static feature data acquired in real time.

Images