An accident judgment and data restoration system based on big data

Abstract

The application discloses an accident judgment and data restoration system based on big data, which comprises a data acquisition module, a data acquisition module reads a large amount of vehicle operation original data in a vehicle networking big data platform, through information collected by a vehicle-mounted device, the information is synchronized to a cloud platform and then enters a data acquisition layer of a data warehouse; a data integration module, a data cleaning module performs standardized processing on the collected original data; an accident identification model is constructed, a deep learning algorithm is used to train and optimize the model; an accident identification design module provides vehicle collision early warning information as a perception device field to enter the model design of accident identification; a high-risk driving segment screening rule design module calls out possible high-risk driving segments through a screening rule and confirms possible accident conditions; a result output module pushes operation data to a platform and performs display analysis, forming an accident report, the application realizes timely identification of various high-risk operation scenes and accidents, greatly improves the identification accuracy and efficiency of traffic accidents.

Description

Technical Field

[0001] This invention belongs to the field of accident assessment technology, specifically relating to an accident assessment and data reconstruction system based on big data. Background Technology

[0002] With the rapid increase in car ownership and the increasing complexity of transportation networks, traffic accidents are frequent, causing enormous loss of life and property. Currently, the main method for obtaining accident data is through customer and driver feedback, with maintenance personnel manually recording relevant accident information. However, this method is susceptible to errors and inaccuracies in manual recording, and the time of the accident reported by the driver is sometimes not the exact time of the accident, making it impossible to accurately pinpoint the exact moment of the accident using only manual accident record sheets.

[0003] Traditional vehicle accident analysis methods primarily rely on eyewitness accounts and accident scene investigations. However, these methods are heavily dependent on individual cognitive biases and memory errors, and dashcams typically only cover a limited field of view, potentially leading to incomplete and inaccurate analyses of accident causes. Furthermore, existing collision detection technologies suffer from false alarms, missed alarms, and insufficient accuracy, and are not compatible with all vehicle models. Their detection algorithms are also simplistic and cover a limited range of collision scenarios. These issues severely impact the iteration speed of vehicle safety-related products and the accuracy of accident analysis.

[0004] To address the aforementioned issues, a system capable of accurately identifying vehicle accidents and building a historical accident case database is urgently needed. This system should be able to analyze accident causes, improve vehicle design and driving behavior, and reduce the accident rate. Simultaneously, how to utilize real-time vehicle operation data collected by onboard equipment to predict operational risks and proactively control factors contributing to these risks, transforming traffic accidents from a reactive response to proactive prevention, is also a pressing issue. Therefore, developing a big data-based accident assessment and data reconstruction system to achieve automated, tiered identification and processing of traffic accidents, thereby improving the efficiency and safety of highway operation and management, is of significant practical importance. Summary of the Invention

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A big data-based accident judgment and data restoration system, comprising: The data acquisition module reads a large amount of raw vehicle operation data from the vehicle network big data platform. It collects basic vehicle information, operation trajectory data, vehicle equipment data and environmental information through on-board equipment. After being synchronized to the cloud platform, the data is then entered into the data acquisition layer of the data warehouse. The data integration module includes a data cleaning module and a feature extraction module. The data cleaning module standardizes the collected raw data; the feature extraction module extracts features of the operation segments based on the vehicle's operating patterns, including speed change features, acceleration features, and position change features. The model training module constructs an accident identification model based on historical accident data and feature data, and uses deep learning algorithms to train and optimize the model; The accident recognition design module provides vehicle collision warning information, which is used as a field of the sensing device in the accident recognition model design; The high-risk driving segment screening rule design module sets high-risk driving segment screening rules, retrieves possible high-risk driving segments based on the screening rules, and then confirms possible accident situations through the model; The results output module includes an accident identification results display module and a data reconstruction module. The accident identification results display module pushes operational data identified by the model as high-risk or accident segments to the platform for display and analysis, identifies the specific time point of the suspected accident, and combines this with video judgment from vehicle-mounted cameras or other manually reported information to ultimately confirm whether it is an actual accident. Based on the data before the accident, the reconstruction module conducts in-depth analysis of the accident causes, reconstructs the accident process, and generates an accident report.

[0006] Preferably, in the accident identification design module, the calculation formula for accident identification is as follows: (1), in, y The probability of an accident is identified. for Vehicle driving parameters, road conditions, and warning factor fields from sensing devices. These represent the corresponding weights for each factor field.

[0007] Preferably, the weights in the accident identification calculation formula are derived from a dataset formed by driving records based on historical vehicle data of the current road segment, using a logistic regression algorithm for model design. The specific model design process includes: Dataset construction: A large number of vehicle trajectory records were collected through in-vehicle equipment; at the same time, accident records were acquired, and the record segments of the five minutes before and the fifteen minutes after the accident time point were centrally organized as the accident occurrence segments; Establish the following linear regression model: (2), in: y_predw0 is the fuel consumption per unit mileage predicted by the model; w1, w2, ..., w are the regression coefficients to be solved, i.e., the weighting factors of each driving behavior feature; x1, x2, ..., x are the standardized driving behavior feature values. Model training and weight calculation: Loss function: The mean squared error is used as the loss function. Loss = (1 / N) Σ ( y_true - y_pred )²(3) Optimization algorithm: Use the least squares (OLS) method or gradient descent to minimize the loss function and solve for the optimal weight vector. W* = [ w1*, w2*, ..., w* ]; Regularization: To prevent overfitting, L2 regularization is introduced. Loss=MSE+λΣwᵢ²) (4), Where λ is the regularization intensity hyperparameter.

[0008] Preferably, the model design process includes: if wi∙*>0: it indicates that the driving behavior is positively correlated with fuel consumption, and the more frequent the behavior, the higher the fuel consumption; if wi∙*<0: it is theoretically unreasonable, indicating that there is a problem with feature engineering or strong collinearity, which needs to be investigated; numerical magnitude: the larger the absolute value of |wi*|, the more significant the impact of the driving behavior on fuel consumption.

[0009] Preferably, the model design process includes: after training, the obtained w*, w1*, w2*, ..., w* are the scientific weight factors of each driving behavior feature; when calculating the deviation score, these regression coefficients are directly used as weight coefficients _i: single feature deviation score _i = |standardized value _i| × |wᵢ*|.

[0010] Preferably, model design also includes model validation and updating. The validation methods include: dividing the dataset into a 70% training set and a 30% test set; calculating the R² coefficient of determination and the MAE metric on the test set to evaluate the model's prediction accuracy; Dynamic automatic update: As new data continues to accumulate, the model is retrained periodically and the weight factors are updated. Scenario-based modeling: Based on the actual results of data collection, further refine the model by constructing independent regression models for different road types and weather conditions to obtain more refined weighting factors.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By reading a large amount of raw vehicle operation data from the vehicle network big data platform, and combining the extraction of dangerous operation segments with the training of the accident identification model, the system enables timely identification of various high-risk operation scenarios and accidents, greatly improving the accuracy and efficiency of traffic accident identification. Furthermore, the system supports the tracing and reconstruction of accident scenes and their causes based on data, which can realistically restore the cause of the accident. Therefore, it overcomes the problems of human recording bias and memory error that are prone to occur in traditional manual recording and driver feedback, as well as illegal and criminal acts such as malicious false reporting to defraud insurance, thus improving the accuracy of accident analysis. 2. This system has established a complete vehicle operation risk assessment mechanism, which can promptly identify and remind drivers of potential safety hazards, effectively reducing the traffic accident rate and increasing the driver alerting capability compared to general methods; 3. Through in-depth mining and intelligent analysis of big data on vehicle operation, this system can automatically identify and process complex collision accident scenarios, improve the iteration speed of vehicle safety-related products and the accuracy of accident analysis, and pre-set simple business model rules for predicting dangerous behavior segments, performing collision accident algorithm prediction only on dangerous behavior segments, reducing the computational load of complex models. 4. This system takes into account the large number of newly installed collision warning devices such as AEB, BSD, and ADAS. Therefore, these fields are introduced to form two types of models: one with collision warning fields and one without collision warning fields. This increases the recognition accuracy of vehicles with collision warning devices. Attached Figure Description

[0012] Figure 1 Design process for accident identification algorithm; Figure 2 This is the actual accident determination process. Detailed Implementation

[0013] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Example

[0014] This invention provides an accident judgment and data reconstruction system based on big data. The system includes data acquisition, data integration, accident identification model design and result output.

[0015] 1. Data Acquisition: This is used to read a large amount of raw vehicle operation data from the Internet of Vehicles big data platform. It collects data such as basic vehicle information, operation trajectory data, vehicle equipment data and environmental information through various vehicle-mounted devices. After being synchronized to the cloud platform, the data is then sent to the data warehouse data acquisition layer.

[0016] 2. Data Integration: This includes a data cleaning module, a feature extraction module, and a model training module. The data cleaning module standardizes the collected raw data, removing outliers and invalid data; the feature extraction module extracts features of the driving segments based on vehicle operation patterns, including speed change features, acceleration features, and position change features; the model training module constructs an accident recognition model based on historical accident data and feature data, and uses deep learning algorithms to train and optimize the model.

[0017] 3. Accident Recognition Design: Since current in-vehicle systems such as AEB, ADAS, and BSD are equipped with vehicle-to-everything (V2X) sensing devices such as radar and cameras, providing vehicle collision warning information is a crucial part of the model design. This information will be incorporated as a field from the sensing devices into the accident recognition model. The calculation formula for accident recognition is as follows: (1), in, y The probability of an accident is identified. for Vehicle driving parameters, road conditions, and warning factor fields from sensing devices. These represent the corresponding weights for each factor field.

[0018] The weights in the specific accident identification calculation formula are derived from a dataset formed by historical vehicle data for the current road segment. The model is designed using a logistic regression algorithm, and the weight factors are obtained as follows: (1) Dataset construction: A large number of vehicle trajectory records are collected through in-vehicle devices such as OBD, AEB, and T-box; at the same time, accident records are obtained from the accident records of insurance companies, coordination companies, or fleets, and the record segments of the five minutes before and the fifteen minutes after the accident time are centrally organized as the accident occurrence segments. If the accident records are insufficient, that is, at least 100 accident records, they can be supplemented by high-risk driving behavior events provided by customer service.

[0019] Sample Definition: We define a sample as a single vehicle and a single time segment. Here, we assume that an accident would necessarily involve the driver slowing down and stopping to check. Therefore, we capture other stopping moments, identifying these stopping points as accident points, and record them as data points indicating no accident-related behavior. In our constructed dataset, accident records account for 5% to 10% of the data.

[0020] Feature variables (X): Define variables around the time of the accident: X1: Average speed 1 minute before stopping; X2: Maximum speed 1 minute before stopping; X3: Duration of deceleration > 2 m / s2 1 minute before stopping; X4: Collision warning duration indicated by AEB or BSD; X5: Stopping time after stopping; X6: Vehicle type; ..., more features can be added as needed.

[0021] Target variable y: The actual accident record corresponding to this sample, 0 indicates no accident, 1 indicates an accident, or a high-risk driving behavior event as determined by the customer.

[0022] The final structured dataset is: D={(X_i, y_i) | i=1,2,...,N}, where N is the total number of samples.

[0023] 2. Model Design: Multiple Linear Regression Establish the following linear regression model: (2), in: y_pred It is the fuel consumption per unit mileage predicted by the model; w 0 The intercept term includes the baseline fuel consumption, which reflects inherent factors such as the vehicle's base fuel consumption and road conditions. w 1 , w 2 , ..., w These are the regression coefficients to be solved, i.e., the weighting factors of each driving behavior feature; x 1 , x 2 , ..., x These are standardized driving behavior feature values. The feature values ​​are first standardized by Z-score on the original variable fields to make the features comparable.

[0024] The model design here will be divided into two categories: 1) Collision warning devices such as AEB and BSD exist and can provide collision warnings. The model feature values ​​will include these collision warning fields. 2) If there are no collision warning devices such as AEB or BSD, and collision warning fields cannot be provided, then the model feature values ​​will not include these collision warning fields. Only vehicle performance, driving behavior, and environmental fields will form the model candidate fields.

[0025] 3. Model training and weight calculation: Loss function: The mean squared error is used as the loss function. Loss = (1 / N) Σ ( y_true - y_pred )²(3) Optimization Algorithm: Use the least squares (OLS) method or gradient descent to minimize the loss function and solve for the optimal weight vector W* = [ w 1 *, w 2 *, ..., w *].

[0026] Regularization: To prevent overfitting, L2 regularization (Ridge regression) can be introduced: Loss=MSE+λΣ wᵢ²) (4), Where λ is the regularization intensity hyperparameter.

[0027] 4. Interpretation and application of weighting factors: Symbol explanation: like w i *>0: This indicates a positive correlation between driving behavior and fuel consumption; the more frequent the behavior, the higher the fuel consumption. If w i *<0: Theoretically unreasonable, possibly indicating a problem with feature engineering or strong multicollinearity, requiring investigation. Numerical value: | w i The larger the absolute value of *|, the more significant the impact of the driving behavior on fuel consumption.

[0028] Practical application: After training, the results are... w 1 *, w 2 *, ..., w *This refers to the scientific weighting factors for each driving behavior characteristic.

[0029] When calculating the deviation score, these regression coefficients are directly used as weight coefficients _i: single feature deviation score _i = |standardized value _i| × |wᵢ*|, and the absolute value is taken to ensure that the deviation score is always positive.

[0030] 5. Model Validation and Update: Validation method: The dataset is divided into a training set (70%) and a test set (30%). R² coefficient of determination and MAE (mean absolute error) are calculated on the test set to evaluate the model's prediction accuracy.

[0031] Dynamic automatic updates: As new data continues to accumulate, the model is retrained periodically and the weight factors are updated to ensure its timeliness and accuracy.

[0032] Scenario-based modeling: Based on the actual results of data collection, independent regression models can be constructed for different road types and weather conditions to obtain more refined weighting factors.

[0033] 4. High-risk driving segment screening rule design module.

[0034] Considering that vehicle data is collected almost every second, resulting in a massive amount of data, we designed high-risk driving segment filtering rules in practical applications to minimize the data volume. These simple filtering rules identify potentially high-risk driving segments, which are then further analyzed using a model to confirm possible accident scenarios. The selection criteria were derived through factor analysis of historical accident data, and the main criteria include, but are not limited to: The emergency braking time in the first 30 seconds is >= 3 seconds; the current speed is 0; the maximum deceleration in the first 30 seconds is > 2.4 m / s; and a rapid change in direction occurs in the first 30 seconds.

[0035] 5. Result Output Module: This module includes an accident identification result display module and a data reconstruction module. The accident identification result display module pushes operational data identified by the model as high-risk or accident segments to the platform for display and analysis, identifying the specific time point of the suspected accident. Based on data from 60 seconds prior to the accident, the reconstruction module conducts in-depth analysis of the accident's cause, reconstructing the accident process, including accident type, cause, and course of events, and generating an accident report.

[0036] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A big data based accident judgment and data restoration system, characterized in that, include: The data acquisition module reads a large amount of raw vehicle operation data from the vehicle network big data platform. It collects basic vehicle information, operation trajectory data, vehicle equipment data and environmental information through on-board equipment. After being synchronized to the cloud platform, the data is then entered into the data acquisition layer of the data warehouse. The data integration module includes a data cleaning module, a feature extraction module, and a model training module. The data cleaning module standardizes the collected raw data; the feature extraction module... The extraction module extracts features of the running segments based on the vehicle's operating patterns, including features of vehicle speed change, acceleration, and position change. The model training module constructs an accident identification model based on historical accident data and feature data, and uses deep learning algorithms to train and optimize the model; The accident recognition design module provides vehicle collision warning information, which is used as a field of the sensing device in the accident recognition model design; The high-risk driving segment screening rule design module sets high-risk driving segment screening rules, retrieves possible high-risk driving segments based on the screening rules, and then confirms possible accident situations through the model; The results output module includes an accident identification results display module and a data restoration module. The accident identification results display module pushes the operational data that the model determines to be high-risk or accident segments to the platform for display and analysis, and identifies the specific time point of the suspected accident. Based on the data before the accident, the restoration module conducts in-depth analysis of the cause of the accident, reconstructs the accident process, and generates an accident report.

2. The accident judgment and data reconstruction system based on big data according to claim 1, characterized in that, In the accident identification design module, the calculation formula for accident identification is as follows: （1）， in, y The probability of an accident is identified. for Vehicle driving parameters, road conditions, and warning factor fields from sensing devices. These represent the corresponding weights for each factor field.

3. The accident judgment and data reconstruction system based on big data according to claim 1, characterized in that, The weights in the accident identification calculation formula are derived from a dataset formed by historical vehicle data for the current road segment. A logistic regression algorithm is used to design the model, and the resulting weight factors are the basis for the model design process. Dataset construction: A large number of vehicle trajectory records were collected through in-vehicle equipment; at the same time, accident records were acquired, and the record segments of the five minutes before and the fifteen minutes after the accident time point were centrally organized as the accident occurrence segments; Establish the following linear regression model: （2）， in: y_pred It is the fuel consumption per unit mileage predicted by the model; w 0 It is the intercept term; w 1 ,w 2 ,...,w These are the regression coefficients to be solved, i.e., the weighting factors of each driving behavior feature; x 1 , x 2 ,..., x These are standardized driving behavior characteristic values; Model training and weight calculation: Loss function: The mean squared error is used as the loss function. Loss = (1 / N) Σ (y _ true - y _ pred)²(3), Optimization algorithm: Use the least squares (OLS) method or gradient descent to minimize the loss function and solve for the optimal weight vector. W* = [ w 1 *, w 2 *, ..., w* ]; Regularization: To prevent overfitting, L2 regularization is introduced. Loss = MSE + λ Σ wᵢ²) (4), Where λ is the regularization intensity hyperparameter.

4. The accident judgment and data reconstruction system based on big data according to claim 3, characterized in that, The model design process includes: If w i • * >0: This indicates a positive correlation between driving behavior and fuel consumption; the more frequent the behavior, the higher the fuel consumption. w i • * <0: Theoretically unreasonable, indicating a problem with feature engineering or strong multicollinearity, requiring investigation; Numerical value: | w i The larger the absolute value of *|, the more significant the impact of the driving behavior on fuel consumption.

5. The accident judgment and data reconstruction system based on big data according to claim 4, characterized in that, The model design process includes: after training, obtaining w*, w 1 * , w 2 * , ... , w* These are the scientific weighting factors for each driving behavior characteristic; when calculating the deviation score, these regression coefficients are directly used as weighting coefficients _i: Single feature deviation score _i = |standardized value _i| × |wᵢ*|.

6. The accident judgment and data reconstruction system based on big data according to claim 4, characterized in that, Model design also includes model validation and updates. The validation methods include: dividing the dataset into a 70% training set and a 30% test set; calculating the R² coefficient of determination and the MAE metric on the test set to evaluate the model's prediction accuracy; Dynamic automatic update: As new data continues to accumulate, the model is retrained periodically and the weight factors are updated. Scenario-based modeling: Based on the actual results of data collection, further refine the model by constructing independent regression models for different road types and weather conditions to obtain more refined weighting factors.

Images