Intelligent traffic accident prediction and emergency response system and method based on vehicle-mounted data

Abstract

The present invention relates to the field of intelligent transportation technology, specifically to an intelligent traffic accident prediction and emergency response system and method based on vehicle-mounted data; the method steps are: collecting vehicle-mounted data, and then using an adaptive convolutional neural network to denoise the vehicle-mounted data; based on the denoised data, building an intelligent traffic accident prediction model, using bidirectional LSTM, attention mechanism, residual connection and Dropout layer to capture key features in time series data and predict the probability of an accident; according to the prediction results, dynamically adjust the warning threshold and trigger different levels of alarms, and take corresponding measures through automatic emergency response or driver reminders; by collecting data before the accident, optimize the prediction model and emergency response strategy, continuously improve the prediction accuracy and enhance traffic management efficiency. The present invention comprehensively utilizes vehicle-mounted data and intelligent algorithms to provide accurate accident prediction and efficient emergency response, and improve traffic safety and traffic efficiency.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to an intelligent traffic accident prediction and emergency response system and method based on vehicle-mounted data. Background Art

[0002] As urban traffic becomes increasingly complex, frequent traffic accidents have become a major challenge in global traffic management. Traditional traffic accident warning and emergency response systems mostly rely on manual observation or simple historical data analysis, which cannot accurately reflect the complex and changeable traffic environment and driving behavior in real time, resulting in poor warning timeliness and low emergency response efficiency. In recent years, the rapid development of vehicle-mounted data acquisition and intelligent analysis technology has provided new solutions for intelligent transportation systems. Vehicle-mounted sensors and V2X communication technology make it possible to obtain vehicle status, environmental information and surrounding traffic conditions in real time, thereby providing more accurate data support for accident prediction and emergency response.

[0003] A traffic accident prediction method is disclosed in a Chinese invention patent application with announcement number CN116110219B, which specifically includes a traffic accident prediction method based on system identification, a traffic accident prediction method based on LSTM, a traffic accident prediction method based on system identification and LSTM residual combination, a traffic accident prediction method based on system identification and LSTM linear weighted combination, and a traffic accident prediction method based on system identification and LSTM pipeline. The traffic accident prediction method based on system identification or the traffic accident prediction method based on LSTM or the above two prediction methods are combined in different ways, and predictions are compared. The results show that the above different methods can improve the prediction ability of road traffic accidents.

[0004] However, existing methods often have problems such as low data processing efficiency, long response time, and inaccurate intelligent decision-making. How to effectively process massive vehicle data and accurately predict traffic accident risks remains an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose an intelligent traffic accident prediction and emergency response system and method based on vehicle-mounted data.

[0006] The technical solution of the present invention is an intelligent traffic accident prediction and emergency response method based on vehicle-mounted data, which includes the following specific implementation steps:

[0007] S1, collect vehicle data;

[0008] S2. The vehicle data denoising method based on adaptive convolutional neural network ACNN standardizes the vehicle data, divides the time window and uses adaptive convolution kernel to filter the noise. After the data is processed by convolution layer, noise detection and activation function, a denoised signal is generated to reconstruct the complete time series data, and a hash function and random selection are combined to generate the batch sequence.

[0009] S3: Review the legality and rationality of vehicle data, construct the moment feature vector, and use the bidirectional long short-term memory network LSTM, attention mechanism, residual connection and Dropout layer to build an intelligent prediction model, dynamically adjust the focus, prevent gradient disappearance and overfitting, and then calculate the probability of traffic accidents through the fully connected layer;

[0010] S4. Dynamically adjust the warning threshold, and automatically trigger different levels of warnings based on the predicted accident probability and the adjusted threshold: severe risk, high risk, medium risk and low risk, and output the warning level;

[0011] S5. After the warning is triggered, emergency measures are automatically or manually initiated according to the situation;

[0012] S6. Continuously optimize prediction accuracy by collecting and analyzing data before an accident occurs, use reinforcement learning technology to adjust accident response strategies, and dynamically adjust traffic flow control strategies.

[0013] Preferably, the implementation process of the vehicle data denoising method based on the adaptive convolutional neural network ACNN is as follows:

[0014] S21, the original vehicle data is X=[x 1 ,x 2 ,…,x n ] is standardized, and the standardization formula is:

[0015] ;

[0016] Where, X norm represents the standardized vehicle data; μ represents the mean of the data set; σ represents the standard deviation of the data set; n represents the total amount of data;

[0017] S22, the vehicle data is divided into multiple time windows according to the timestamp, and the time series data in each window will be used as the input of the convolutional neural network. The time series data is D(t)=[d 1 ,d 2 ,…,d t ], dividing it into time windows of length w 1 , Window 2 ,…,Window i ,…,Windowk ;

[0018] Among them, Window i Represents the i-th time window: Window i =[d i ,d i+1 ,…,d i+w-1 ], i=1,2,…,k; k represents the number of divided time windows;

[0019] S23, construct a data denoising model based on a convolutional neural network with adaptive filtering capabilities, denoise the data in any divided time window, and output the denoised signal ;

[0020] S24, denoising data of all time windows Splice and reconstruct complete time series data : , output the processed vehicle data ;

[0021] in, represents the denoised data in the i-th time window.

[0022] Preferably, the denoising process of the data denoising model based on the convolutional neural network with adaptive filtering capability is as follows:

[0023] S31. Convolution layer design: The convolution layer of ACNN uses an adaptive convolution kernel. The convolution operation can extract local features in the data and remove high-frequency noise. The convolution operation formula is:

[0024] ;

[0025] Where X(t+i) represents the data at the t+ith moment of the input data; W i represents the i-th weight of the convolution kernel; Y(t) represents the output signal obtained after convolution;

[0026] S32, Adaptive convolution kernel adjustment: Construct an adaptive convolution kernel based on gradient descent optimization, and the weight of the adaptive convolution kernel is W i Adjusted by back-propagation according to the noise characteristics of the input data;

[0027] Define the objective function as L(W), and the loss function as the output signal Y(t) and the true signal Y true Differences in (t):

[0028] ;

[0029] Where Y true(t) represents the real noise-free data; λ is the regularization term used to prevent overfitting;

[0030] Based on this, the convolution kernel most suitable for denoising is trained by minimizing the loss function;

[0031] S33, activation function and nonlinear mapping: The signal Y(t) output by the convolutional layer is nonlinearly mapped by the activation function ReLU to obtain the output signal Y out (t):

[0032] Y out (t)=max(0,Y(t));

[0033] S34, Noise detection: During the training process of the convolutional neural network, a noise detection unit is added to evaluate the noise intensity of each time window. The noise intensity N(t) is calculated as follows:

[0034] N(t)=|X(t)-Yout(t)|;

[0035] Where X(t) represents the input vehicle data;

[0036] If N(t) is greater than the set threshold θ, the data point is considered to be noise and needs further processing;

[0037] S35, denoising output: The final denoising output data is the denoised signal generated by the convolutional neural network Combined with the correction result of the noise detection unit, we can get:

[0038] ;

[0039] In the formula, Represents the denoised data. If the noise detection value N(t) is lower than the threshold θ, the denoised signal is output; otherwise, the original data remains unchanged.

[0040] Preferably, the generation process of the sequence to be approved is as follows:

[0041] S41, converting the processed vehicle data into a binary string sequence data;

[0042] S42, calculating auxiliary sequence generation parameter τ=H(data);

[0043] Where H is a predefined hash function, {0,1}→ ; is the ring of integers modulo η; η=p×q; p and q are predefined prime numbers;

[0044] S43, randomly select a bit b∈{0,1} and calculate the sequence Sa to be approved:

[0045] If b=0, Sa=τ ε mod η;

[0046] If b=1, Sa=(υτ) ε mod η;

[0047] Where ε is a predefined first-order sequence generator code, ε×e mod [(p-1)×(q-1)]; e is a predefined auxiliary sequence generator code, which is a prime number that satisfies that e and η are mutually prime; υ ​​represents a predefined second-order sequence generator code, υ∈ .

[0048] Preferably, the review process for the legality and rationality of the vehicle data is as follows:

[0049] S51, converting the vehicle-borne data into a binary string sequence data';

[0050] S52, calculating auxiliary sequence analysis parameter τ'=H(data');

[0051] S53. The inspection conditions are as follows:

[0052] Condition 1: (Sa) e =τ' mod η;

[0053] Condition 2: (Sa) e =υ×τ' mod η;

[0054] In the formula, η=p×q; p and q are predefined prime numbers; υ represents a predefined second-order sequence generator code; H is a predefined hash function; Sa represents the sequence to be approved;

[0055] If condition 1 or condition 2 is met, the vehicle data is considered The rationality and accuracy of the vehicle data after approval Extract key features.

[0056] Preferably, the prediction process of the intelligent prediction model is as follows:

[0057] S61, capture the dependencies in the time series data, use bidirectional LSTM to process the input data, and bidirectional LSTM transmits data in both forward and backward directions simultaneously;

[0058] The output of the bidirectional LSTM is: ;

[0059] in, Represents the output of the forward LSTM; Represents the output of backward LSTM;

[0060] S62. Apply the attention mechanism to the output of the bidirectional LSTM and dynamically adjust the model's attention to the features at each moment by calculating a weight coefficient for each time step.

[0061] Define the output of bidirectional LSTM as h t , attention weight α t The calculation formula is as follows:

[0062] ;

[0063] In the formula, α t represents the attention weight of the current time step t; h t represents the hidden state at time step t; w represents the weight vector, that is, the importance of time t in the overall time series; w T represents the transposed vector of the weight vector; t' represents the index of all time steps;

[0064] S63, introduce residual connection, and convert the current LSTM output h t The output of the previous moment After addition, the output of the residual connection is: ;

[0065] In the formula, Represents the output of the residual connection;

[0066] S64. Apply the Dropout layer to the output of each layer. Dropout forces the model to learn more robust feature representation by randomly discarding some neurons, and obtains the output , define the Dropout rate as p, then the operation of the Dropout layer is expressed as:

[0067] ;

[0068] The output of S65 and Dropout layers will be input to the fully connected layer, which will further process the features through weighting and biasing, and finally obtain the probability of traffic accidents at each moment P. accident (t), the output layer is calculated as:

[0069] ;

[0070] Where W f represents the weight matrix of the fully connected layer, which is used to weight the hidden state of the input; b f Represents the bias term, which is used to adjust the output of the model.

[0071] Preferably, the adjustment process of dynamically adjusting the warning threshold is:

[0072] ;

[0073] Where θ(t) represents the dynamically adjusted warning threshold; θ fixed represents the set initial fixed threshold; α represents the weighting coefficient of historical accident data; represents the change in the historical accident rate; β represents the traffic density weighting coefficient; D traffic (t) represents the current traffic density; γ represents the weighting coefficient of weather factors; W weather (t) indicates the current weather conditions.

[0074] Preferably, the warning level classification rules are:

[0075] Red warning, i.e., serious accident risk: When the predicted probability exceeds 90% of the dynamic threshold θ(t), a red warning is triggered, indicating a very high accident risk and requiring immediate emergency response measures: P accident (t)>0.9×θ(t);

[0076] Orange warning, i.e. higher accident risk: When the predicted probability is between 80% and 90%, an orange warning is triggered, indicating a higher accident risk: 0.8×θ(t) ≤P accident (t)≤0.9×θ(t);

[0077] Yellow warning, i.e. medium accident risk: When the predicted probability is between 60% and 80%, a yellow warning is triggered, prompting the driver to drive cautiously and there is a certain accident risk: 0.6×θ(t) ≤P accident (t)<0.8×θ(t);

[0078] Green warning, which means low accident risk: When the predicted probability is less than 60%, it means the probability of an accident is low, indicating smooth traffic: P accident (t)<0.6×θ(t).

[0079] The technical solution of the present invention is: an intelligent traffic accident prediction and emergency response system based on vehicle-mounted data, which is used to execute the above-mentioned intelligent traffic accident prediction and emergency response method based on vehicle-mounted data, comprising:

[0080] The vehicle data acquisition module is used to collect vehicle driving data through vehicle-mounted equipment and obtain environmental data through vehicle wireless communication technology V2X communication;

[0081] On-board data processing module, used to improve data quality by using data pre-processing technology;

[0082] Intelligent traffic accident prediction module, which is used to build an intelligent prediction model, conduct accident risk assessment on current traffic conditions, and calculate the probability of accidents occurring in the future;

[0083] Intelligent early warning and risk warning module, used to generate early warning information based on the risk assessment results of the intelligent prediction model;

[0084] The emergency response dispatch module is used to automatically or manually initiate emergency measures according to the situation after the warning is triggered;

[0085] The traffic accident data analysis and optimization module is used to continuously optimize the prediction accuracy by collecting and analyzing data before the accident occurs, and use reinforcement learning technology to adjust the accident response strategy and dynamically adjust the traffic flow control strategy.

[0086] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0087] The present invention designs an intelligent traffic accident prediction and emergency response system and method based on vehicle-mounted data, which realizes efficient and accurate traffic accident prediction and intelligent emergency response by fully integrating vehicle-mounted data, external environment information and advanced deep learning algorithms:

[0088] (1) Improving the accuracy of accident prediction: The adaptive filtering-based convolutional neural network (ACNN) is used to denoise the vehicle data, which effectively improves the quality of the vehicle data and the accuracy of the prediction model. By extracting multi-dimensional features such as vehicle dynamic behavior, environmental information, and driver operation behavior, combined with bidirectional LSTM and attention mechanism, the model can fully capture the complex relationships in time series data, thereby improving the accuracy of traffic accident prediction;

[0089] (2) Dynamic threshold adjustment improves warning sensitivity: A dynamic threshold adjustment mechanism is introduced to flexibly adjust the warning threshold based on historical accident data, real-time traffic density, weather and road conditions. This mechanism makes the warning system more adaptable to different traffic environments, improves the accuracy of warnings and the timeliness of responses, optimizes the timing of warning triggering under different conditions, prevents false alarms and missed alarms, and enhances the intelligence of the system.

[0090] (3) Efficient data processing and model training: The deep learning model that combines convolutional neural network (CNN) and bidirectional LSTM (BiLSTM) can efficiently capture the complex dependencies in time series data. The residual connection and dropout layer in the processing process effectively avoid overfitting and improve the generalization ability of the model.

[0091] (4) Optimizing traffic management and decision support: The present invention not only improves traffic safety through accident prediction and early warning, but also optimizes traffic management strategies through analysis of traffic accident data. By analyzing and reinforcing data before an accident occurs, the present invention can continuously optimize accident response strategies, identify accident-prone areas, and provide decision support for traffic management departments, thereby optimizing traffic control measures such as traffic light signal control and speed limit management, thereby improving traffic safety and traffic efficiency.

[0092] (5) Intelligent emergency response mechanism: The present invention combines an intelligent traffic management system with automatic emergency response and driver reminders to take corresponding emergency measures under different levels of warnings. For example, a red warning triggers an automatic emergency braking system, and orange and yellow warnings remind drivers to adopt a cautious driving strategy through voice or visual interfaces. This automated and intelligent response can effectively reduce the occurrence of accidents and mitigate the losses caused by accidents.

[0093] (6) Generation and approval of the sequence to be approved: The vehicle data is converted into a binary sequence, and a unique identifier is generated using a hash function. The sequence to be approved is generated in combination with random bits to ensure the accuracy of the subsequent model input data. Modulo operations and predefined sequence generation codes are used to ensure data integrity and security. Finally, the sequence to be approved is transmitted together with the hash value to the intelligent traffic accident prediction module for data verification and subsequent processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 This is a system structure diagram of an intelligent traffic accident prediction and emergency response system based on vehicle-mounted data proposed by the present invention;

[0095] Figure 2 This is a method flow chart of an intelligent traffic accident prediction and emergency response method based on vehicle-mounted data proposed by the present invention. DETAILED DESCRIPTION

[0096] Embodiment 1, as Figure 1 As shown, the present invention proposes an intelligent traffic accident prediction and emergency response system based on vehicle-mounted data, including: an on-board data acquisition module, a data processing and feature extraction module, an intelligent traffic accident prediction module, an intelligent early warning and risk prompting module, an emergency response scheduling module and a traffic accident data analysis and optimization module.

[0097] The vehicle data acquisition module collects vehicle driving data through vehicle-mounted devices {including but not limited to OBD (On-Board Diagnostics), GPS, camera, millimeter-wave radar, laser radar}, including but not limited to speed, acceleration, braking status, steering angle, surrounding environment information, and obtains environmental data through V2X (Vehicle to Everything, vehicle wireless communication technology) communication, including but not limited to surrounding vehicles, traffic signals, and road conditions;

[0098] The on-board data processing module uses data preprocessing technology, including but not limited to data cleaning, denoising, and standardization, to improve data quality;

[0099] The intelligent traffic accident prediction module uses deep learning models (including but not limited to LSTM, gated recurrent unit GRU, Transformer) to train accident prediction models, conduct accident risk assessment on current traffic conditions, and calculate the probability of accidents occurring in the future.

[0100] The intelligent warning and risk reminder module provides real-time risk reminders to drivers (including but not limited to dashboard warnings, HUD projection, and voice reminders) based on the risk assessment results of the accident prediction model, and sends warning information to nearby vehicles through V2X technology to alert other drivers to possible dangerous situations;

[0101] After an accident occurs, the emergency response dispatch module automatically assesses the severity of the accident and sends real-time data to the traffic management center. In combination with the city traffic dispatch system, it optimizes the dispatch path of emergency resources (including but not limited to ambulances, police cars, and fire trucks), shortens the response time, and interacts with the driver through the vehicle communication terminal to provide the best self-rescue plan, including but not limited to guiding vehicles with minor accidents to leave the main road and providing remote medical advice;

[0102] The traffic accident data analysis and optimization module collects historical accident data, builds an accident hotspot analysis model, identifies high-risk sections, and combines reinforcement learning to optimize traffic control strategies, including but not limited to traffic light adjustments and road speed limit adjustments, to reduce the accident rate.

[0103] Embodiment 2, as Figure 2 As shown, the intelligent traffic accident prediction and emergency response method based on vehicle-mounted data proposed in the present invention is applied to the intelligent traffic accident prediction and emergency response system based on vehicle-mounted data proposed in Example 1, and its specific implementation steps are as follows:

[0104] S1. The vehicle data collection module collects vehicle data, including but not limited to:

[0105] S11. Real-time collection of vehicle status data through vehicle-mounted sensors (including but not limited to vehicle-mounted automatic diagnostic system OBD, GPS, inertial measurement unit IMU, camera, millimeter wave radar, laser radar), including but not limited to:

[0106] Vehicle operating status (including but not limited to vehicle speed, throttle opening, braking intensity, steering angle);

[0107] Vehicle dynamics data (including but not limited to acceleration, angular velocity, and yaw rate);

[0108] Vehicle external environment data (including but not limited to obstacles ahead, positions of surrounding vehicles, road signs, and weather conditions);

[0109] S12. Use V2X (Vehicle-to-Everything) communication technology to obtain external data from roadside units (RSU), other vehicles (V2V), and traffic signals (V2I), including but not limited to:

[0110] Traffic congestion and traffic accident information ahead;

[0111] Real-time driving status of other vehicles for traffic flow prediction and collision risk assessment;

[0112] Information on changes to road infrastructure (including but not limited to temporary construction and traffic light status).

[0113] S2, the vehicle data processing module builds a data denoising model based on a convolutional neural network with adaptive filtering capabilities (ACNN). It uses the local perception and adaptive learning capabilities of the convolutional neural network to effectively remove noise from the vehicle data through a deep learning model, while retaining the key features of the data. It adjusts itself according to different types of noise to adapt to the changing vehicle environment. The specific implementation process is as follows:

[0114] S21. Standardize the original vehicle data to ensure that different features have the same dimension. Assume that the original vehicle data is X=[x 1 ,x 2 ,…,x n ], the standardized formula is:

[0115] ;

[0116] Where, X norm represents the standardized vehicle data; μ represents the mean of the data set; σ represents the standard deviation of the data set;

[0117] The mean of the standardized data is 0 and the standard deviation is 1, which helps to speed up model training and improve denoising effect;

[0118] S22, the vehicle data is divided into multiple time windows according to the timestamp, and the time series data in each window will be used as the input of the convolutional neural network. Assume that the time series data is D(t)=[d 1 ,d 2 ,…,d t ], dividing it into time windows of length w 1 , Window 2 ,…,Window i ,…,Window k ;

[0119] Among them, Window i Represents the i-th time window: Window i =[d i ,d i+1 ,…,d i+w-1 ], i=1,2,…,k; k represents the number of divided time windows;

[0120] It should be noted that each time window W i Represents continuous time series data, which is used as input for convolution processing;

[0121] The core of S23 and ACNN is the adaptive filter, which automatically adjusts the filtering capacity according to the noise characteristics of the input data to optimize the denoising effect:

[0122] S2301. Convolution layer design: The convolution layer of ACNN uses an adaptive convolution kernel. The convolution operation can extract local features in the data and remove high-frequency noise. The convolution operation formula is:

[0123] ;

[0124] Where X(t+i) represents the data at the t+ith moment of the input data; W i represents the i-th weight of the convolution kernel; Y(t) represents the output signal obtained after convolution;

[0125] S2302, adaptive convolution kernel adjustment: In order to enable the convolution network to automatically adjust the filtering effect according to different types of data noise, an adaptive convolution kernel based on gradient descent optimization is constructed. The weight W of the adaptive convolution kernel is i Adjusted by back-propagation according to the noise characteristics of the input data;

[0126] Assume that the objective function is L(W), and the loss function is defined as the output signal Y(t) and the true signal Y true Differences in (t):

[0127] ;

[0128] Where Y true (t) represents the real noise-free data; λ is the regularization term used to prevent overfitting;

[0129] Based on this, the convolution kernel most suitable for denoising is trained by minimizing the loss function;

[0130] S2303, activation function and nonlinear mapping: The signal Y(t) output by the convolutional layer is nonlinearly mapped through the activation function (ReLU) to obtain the output signal Y out (t), increase the nonlinear expression ability of the model:

[0131] Y out (t)=max(0,Y(t));

[0132] S2304, noise detection: During the training process of the convolutional neural network, a noise detection unit is added to evaluate the noise intensity of each time window. The noise intensity N(t) is calculated in the following way:

[0133] N(t)=|X(t)-Yout(t)|;

[0134] Where X(t) represents the input vehicle data;

[0135] If N(t) is greater than the set threshold θ, the data point is considered to be noise and needs further processing;

[0136] S2305, denoising output: The final denoising output data is a denoised signal generated by a convolutional neural network Combined with the correction result of the noise detection unit, we can get:

[0137] ;

[0138] In the formula, Represents the denoised data. If the noise detection value N(t) is lower than the threshold θ, the denoised signal is output; otherwise, the original data remains unchanged.

[0139] S24, data reconstruction: splice the denoised data of all time windows to reconstruct the complete time series data : , output the processed vehicle data ;

[0140] in, represents the denoised data in the i-th time window;

[0141] S25, the processed vehicle data Generate the sequence to be approved. The generation process is as follows:

[0142] S2501, the processed vehicle data Convert to binary string sequence data;

[0143] S2502, calculating auxiliary sequence generation parameter τ=H(data);

[0144] Where H is a predefined hash function, {0,1}→ ; is the ring of integers modulo η; η=p×q; p and q are predefined prime numbers;

[0145] S2503, randomly select a bit b∈{0,1} and calculate the sequence to be approved Sa:

[0146] If b=0, Sa=τ ε mod η;

[0147] If b=1, Sa=(υτ) ε mod η;

[0148] Where ε is a predefined first-order sequence generator code, ε×e mod [(p-1)×(q-1)]; e is a predefined auxiliary sequence generator code, which is a prime number that satisfies that e and η are mutually prime; υ ​​represents a predefined second-order sequence generator code, υ∈ ;

[0149] S26. {Sa, }Transmitted to the intelligent traffic accident prediction module.

[0150] S3, the intelligent traffic accident prediction module uses the denoised time series data to predict accidents. By modeling the time series characteristics of the vehicle data, combined with the traffic conditions and driving behavior characteristics, it predicts the risk of traffic accidents in the future. The specific implementation steps are as follows:

[0151] S31, to ensure the accuracy of the features and vehicle data, based on the sequence Sa to be approved, check the received processed vehicle data The rationality and legality of the inspection process is as follows:

[0152] S3101, receive the vehicle data Convert to binary string sequence data';

[0153] S3102, calculating auxiliary sequence analysis parameter τ'=H(data');

[0154] S3103. Inspection conditions are as follows:

[0155] Condition 1: (Sa) e =τ' mod η;

[0156] Condition 2: (Sa) e =υ×τ' mod η;

[0157] In the formula, η=p×q; p and q are predefined prime numbers; υ represents a predefined second-order sequence generator code; H is a predefined hash function; Sa represents the sequence to be approved;

[0158] If condition 1 or condition 2 is met, the received processed vehicle data is considered The rationality and accuracy of the vehicle data after approval Extract key features from

[0159] S32. From the denoised time series data The key features are extracted from the vehicle, including but not limited to the dynamic behavior of the vehicle, environmental information and the driver's operation behavior. The feature vector at each time t can be expressed as X(t), which contains the following dimensions:

[0160] Vehicle behavior characteristics (A(t)): including but not limited to acceleration, speed, braking force, and accelerator pedal pressure. These characteristics reflect the driving state of the vehicle and can indicate whether the vehicle is in a dangerous driving state;

[0161] Environmental characteristics (E(t)): including but not limited to weather conditions, road conditions, and external factors of traffic flow, which directly affect the probability of traffic accidents;

[0162] Driving behavior characteristics (D(t)): including the driver's behavior patterns, including but not limited to sudden acceleration, frequent lane changes, and speeding, which may be potential factors for accidents;

[0163] Based on this: combine vehicle behavior characteristics (A(t)), environmental characteristics (E(t)) and driving behavior characteristics (D(t)) to obtain the feature vector X(t) at each time t;

[0164] S33. Build an intelligent traffic accident prediction model. By combining bidirectional LSTM, attention mechanism, residual connection and Dropout layer, it can capture the complex features in time series data, solve the gradient vanishing problem, increase attention to key time points, and prevent overfitting. The specific implementation process is as follows:

[0165] S3301. In order to more comprehensively capture the dependencies in time series data, a bidirectional LSTM (BiLSTM) is used to process the input data. The bidirectional LSTM transmits data in both the forward and backward directions simultaneously, so that the information flow in the past and the future can be considered at the same time.

[0166] The output of the bidirectional LSTM is: ;

[0167] in, Represents the output of the forward LSTM; Represents the output of backward LSTM;

[0168] S3302. Apply an attention mechanism to the output of the bidirectional LSTM so that the model can focus on certain key time steps. For example, in a complex traffic environment, certain moments may be more critical for predicting traffic accidents. The attention mechanism dynamically adjusts the model's attention to the features at each moment by calculating a weight coefficient for each time step.

[0169] Specifically: Assume that the output of the bidirectional LSTM is h t , attention weight α t The calculation formula is as follows:

[0170] ;

[0171] In the formula, α t represents the attention weight of the current time step t; h t represents the hidden state at time step t; w represents the learned weight vector, that is, the importance of time t in the overall time series; w T represents the transposed vector of the weight vector; t' represents the index of all time steps;

[0172] It should be noted that by introducing the attention mechanism, the weight of attention to forward and backward information can be dynamically adjusted according to the characteristics of the current moment, giving more attention to important moments, thereby improving the model's sensitivity to key moments in time series data. It is not just about processing bidirectional information, but about selecting which information is more valuable to the prediction results;

[0173] S3303, in order to avoid the gradient vanishing problem in deep networks, a residual connection is introduced. The residual connection converts the LSTM output h at the current moment t The output of the previous moment Addition is performed to keep the information flowing and promote stable training of the model;

[0174] Specifically: The output of the residual connection is: ;

[0175] In the formula, Represents the output of the residual connection;

[0176] It should be noted that residual connections allow information to jump between network layers, avoiding the problem of gradient vanishing and information loss, especially in the process of long sequence learning, which can effectively maintain information flow; by adding the output of the current moment to the output of the previous moment, residual connections help the model retain more useful information, thereby improving the training effect and prediction accuracy of the model;

[0177] S3304. In order to prevent overfitting and improve the generalization ability of the model, we apply the Dropout layer to the output of each layer. Dropout forces the model to learn more robust feature representation by randomly discarding some neurons, and obtains the output , assuming the Dropout rate is p, the operation of the Dropout layer can be expressed as:

[0178] ;

[0179] It should be noted that Dropout(p) is a random process that discards inputs with probability p. Specifically, for a neuron i, if it is discarded, Dropout(p)=0, otherwise Dropout(p)=1. The Dropout layer does not simply randomly discard neurons, but uses dynamic Dropout, which means that at the output of each layer, the Dropout discard rate can be dynamically adjusted according to the characteristics of the layer to better adapt to the complexity of information between different layers. This dynamic discarding mechanism can more accurately avoid overfitting and improve the generalization ability of the model, especially in complex and changeable tasks such as traffic accident prediction.

[0180] S3305. The output of the Dropout layer will be input to the fully connected layer, which will further process the features through weighting and biasing, and finally obtain the probability of traffic accidents at each moment P. accident (t), the output layer is calculated as:

[0181] ;

[0182] Where W f represents the weight matrix of the fully connected layer, which is used to weight the hidden state of the input; b f Represents the bias term, which is used to adjust the output of the model.

[0183] S4. Intelligent warning and risk warning module is based on dynamic threshold adjustment of predicted accident probability, real-time multi-level warning response and automatic emergency measures, which can provide real-time and effective accident warning for drivers or traffic management departments, and combine external environmental data (including but not limited to traffic flow, weather information) and intelligent feedback mechanism to more accurately predict accident risks and take timely response measures. The specific implementation process is as follows:

[0184] S41. In order to improve the accuracy and adaptability of warning, we introduce a dynamic threshold adjustment mechanism. The dynamically adjusted threshold θ(t) changes based on historical data and current traffic environment.

[0185] Specifically, consider the following factors to adjust the threshold:

[0186] Historical accident data: If there have been multiple accidents recently, the warning threshold will be raised accordingly, so that earlier warnings can be issued during high-risk periods;

[0187] Real-time traffic density: In areas with dense traffic or road construction, the warning threshold is appropriately lowered and an alarm is issued in advance to prevent accidents;

[0188] Weather and road condition information: In bad weather or slippery road conditions, the warning threshold can be appropriately lowered to enhance sensitivity to potential accidents;

[0189] Specifically, the dynamic threshold adjustment rule can be described by the following formula:

[0190] ;

[0191] In the formula, θ fixed represents the set initial fixed threshold; α represents the weighting coefficient of historical accident data; represents the change in the historical accident rate; β represents the traffic density weighting coefficient; D traffic (t) represents the current traffic density (including but not limited to the number of vehicles or the degree of road congestion); γ represents the weighting coefficient of weather factors; W weather (t) indicates the current weather conditions, including the classification of weather impacts (including but not limited to heavy rain, haze);

[0192] S42. Different levels of alarm triggering are performed based on the predicted accident probability and dynamic threshold. The warning levels can be divided into the following levels:

[0193] Red warning (serious accident risk): When the predicted probability exceeds 90% of the dynamic threshold θ(t), a red warning is triggered, indicating a very high accident risk and requiring immediate emergency response measures (including but not limited to automatic braking and dispatching traffic police): P accident (t)>0.9×θ(t);

[0194] Orange warning (higher accident risk): When the predicted probability is between 80% and 90%, an orange warning is triggered, indicating a higher accident risk, and the driver is advised to remain vigilant and prepare for potential dangers: 0.8×θ(t) ≤P accident (t)≤0.9×θ(t);

[0195] Yellow warning (medium accident risk): When the predicted probability is between 60% and 80%, a yellow warning is triggered, prompting the driver to drive carefully. There may be a certain risk of an accident: 0.6×θ(t) ≤P accident (t)<0.8×θ(t);

[0196] Green warning (low accident risk): When the predicted probability is less than 60%, it means the probability of an accident is low, the system maintains normal operation, and traffic is smooth: P accident (t)<0.6×θ(t);

[0197] S43: Output the warning level result, and transmit the warning level result to the emergency response dispatch module.

[0198] S5. After the accident prediction triggers the warning, the emergency response dispatch module automatically or manually initiates the emergency response measures according to the actual situation. The goal of the warning response is to reduce the occurrence of potential accidents and help drivers and traffic management departments take necessary preventive measures:

[0199] Automatic emergency response: If the warning is a red warning, the first-level emergency procedure will be initiated, including but not limited to automatically activating the emergency braking system and automatically navigating around congested roads;

[0200] Driver reminder: When the warning is orange or yellow, the secondary emergency procedure is activated to remind the driver to slow down or adopt a cautious driving strategy through voice or visual interface.

[0201] S6, traffic accident data analysis and optimization module optimizes the intelligent traffic accident prediction model by collecting data before the accident, continuously improves the prediction accuracy, and uses reinforcement learning technology to adjust and optimize the accident response strategy, thereby improving the efficiency of emergency dispatch;

[0202] Based on this: Through long-term accumulated data, the intelligent traffic accident prediction model can identify the impact of different traffic conditions and environmental factors on the occurrence of accidents, gradually learn and optimize its prediction ability, so as to make more accurate judgments in future traffic environments, and at the same time, by analyzing traffic management data, identify high-incidence areas of accidents, and then take corresponding measures, including but not limited to optimizing traffic light signal control and speed limit management to reduce the possibility of accidents. In addition, combined with advanced road infrastructure such as smart street lights and variable speed limit signs, traffic flow control strategies are dynamically adjusted according to real-time traffic flow, weather conditions and other factors, thereby further improving traffic safety and traffic efficiency.

[0203] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. An intelligent traffic accident prediction and emergency response method based on vehicle-mounted data, characterized in that: The specific implementation steps include the following: S1, collect vehicle data; S2. The vehicle data denoising method based on adaptive convolutional neural network ACNN standardizes the vehicle data, divides the time window and uses adaptive convolution kernel to filter the noise. After the data is processed by convolution layer, noise detection and activation function, a denoised signal is generated to reconstruct the complete time series data, and a hash function and random selection are combined to generate the batch sequence. The process of generating the sequence to be batched is as follows: S2501, converting the processed vehicle data into a binary string sequence data; S2502, calculating auxiliary sequence generation parameter τ=H(data); Where H is a predefined hash function, {0,1} * → ; is the ring of integers modulo η; η = p × q; p and q are predefined prime numbers; S2503, randomly select a bit b∈{0,1} and calculate the sequence to be approved Sa: If b=0, Sa=τ ε mod n; If b=1, Sa=(yt) ε mod n; Where ε is a predefined first-order sequence generator code, ε=e mod [(p-1)×(q-1)]; e is a predefined auxiliary sequence generator code, which is a prime number that satisfies that e and η are mutually prime; υ ​​represents a predefined second-order sequence generator code, υ∈ ; S3: The legality and rationality of the vehicle data after approval are checked and processed, and the moment feature vector is constructed. The intelligent prediction model is constructed using the bidirectional long short-term memory network LSTM, attention mechanism, residual connection and Dropout layer to dynamically adjust the focus to prevent gradient disappearance and overfitting, and then the probability of traffic accidents is calculated through the fully connected layer; The review process for the legality and rationality of the vehicle data after approval is as follows: S3101, converting the processed vehicle data into a binary string sequence data'; S3102, calculating auxiliary sequence analysis parameter τ'=H(data'); S3103, the inspection conditions are as follows: Condition 1: (Sa) e =τ' mod η; Condition 2: (Sa) e =υ×τ' mod η; In the formula, η=p×q; p and q are predefined prime numbers; υ represents a predefined second-order sequence generator code; H is a predefined hash function; Sa represents the sequence to be approved; If condition 1 or condition 2 is met, the processed vehicle data is considered It is reasonable and legal, allowing key features to be extracted from the approved processed vehicle data; S4. Dynamically adjust the warning threshold, and automatically trigger different levels of warnings based on the predicted accident probability and the adjusted threshold: severe risk, high risk, medium risk and low risk, and output the warning level; S5. After the warning is triggered, emergency measures are automatically or manually initiated according to the situation; S6. Continuously optimize prediction accuracy by collecting and analyzing data before an accident occurs, use reinforcement learning technology to adjust accident response strategies, and dynamically adjust traffic flow control strategies.

2. The intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to claim 1 is characterized in that: The implementation process of the vehicle data denoising method based on the adaptive convolutional neural network ACNN is as follows: S21, the original vehicle data is X=[x1,x2,…,x n ] is standardized, and the standardization formula is: ； Where, X norm represents the standardized vehicle data; μ represents the mean of the data set; σ represents the standard deviation of the data set; n represents the total amount of data; S22, the vehicle data is divided into multiple time windows according to the timestamp. The time series data in each window will be used as the input of the convolutional neural network. The time series data is D(t)=[d1,d2,…,d t ], which is divided into time windows of length w: Window1, Window2, …, Window i ,…,Window k ; Among them, Window i Represents the i-th time window: Window i =[d i ,d i+1 ,…,d i+w-1 ], i=1,2,…,k; k represents the number of divided time windows; S23, construct a data denoising model based on a convolutional neural network with adaptive filtering capabilities, denoise the data in any divided time window, and output the denoised signal ; S24, denoising data of all time windows Splice and reconstruct complete time series data : , output the processed vehicle data ; in, represents the denoised data in the i-th time window.

3. The intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to claim 2 is characterized in that: The denoising process of the data denoising model based on the convolutional neural network with adaptive filtering capability is as follows: S2301. Convolution layer design: The convolution layer of ACNN uses an adaptive convolution kernel. The convolution operation can extract local features in the data and remove high-frequency noise. The convolution operation formula is: ; Where X(t+i) represents the data at the t+ith moment of the input data; W i represents the i-th weight of the convolution kernel; Y(t) represents the output signal obtained after convolution; S2302, adaptive convolution kernel adjustment: construct an adaptive convolution kernel based on gradient descent optimization, and the weight of the adaptive convolution kernel is W i Adjusted by back-propagation according to the noise characteristics of the input data; Define the objective function as L(W), and the loss function as the output signal Y(t) and the true signal Y true Differences in (t): ; Where Y true (t) represents the real noise-free data; λ is the regularization term used to prevent overfitting; Based on this, the convolution kernel most suitable for denoising is trained by minimizing the loss function; S2303, activation function and nonlinear mapping: The signal Y(t) output by the convolution layer is nonlinearly mapped through the activation function ReLU to obtain the output signal Y out (t), that is, Y out (t)=max(0,Y(t)); S2304, noise detection: During the training process of the convolutional neural network, a noise detection unit is added to evaluate the noise intensity of each time window. The noise intensity N(t) is calculated in the following way: N(t)=|X(t)-Y out (t)|; Where X(t) represents the input vehicle data; If N(t) is greater than the set threshold θ, the data point is considered to be noise and needs further processing; S2305, denoising output: The final denoising output data is a denoised signal generated by a convolutional neural network Combined with the correction result of the noise detection unit, we can get: ； In the formula, Represents the denoised signal. If the noise detection value N(t) is lower than the threshold θ, the denoised signal is output; otherwise, the original data remains unchanged.

4. The intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to claim 1 is characterized in that: The prediction process of the intelligent prediction model is as follows: S3301, capture the dependencies in the time series data, use the bidirectional LSTM to process the input data, and the bidirectional LSTM transmits data in both the forward and backward directions simultaneously; The output of the bidirectional LSTM is: ; in, Represents the output of the forward LSTM; Represents the output of backward LSTM; S3302, applying the attention mechanism to the output of the bidirectional LSTM, dynamically adjusting the model's attention to the features at each moment by calculating a weight coefficient for each time step; Define the output of bidirectional LSTM as h t , attention weight α t The calculation formula is as follows: ； In the formula, α t represents the attention weight of the current time step t; h t represents the hidden state at time step t; w represents the weight vector, that is, the importance of time t in the overall time series; w T represents the transposed vector of the weight vector; t' represents the index of all time steps; S3303, introduce residual connection, and convert the current LSTM output h t The output h of the previous moment t-1 After addition, the output of the residual connection is: ; In the formula, Represents the output of the residual connection; S3304, apply the Dropout layer to the output of each layer. Dropout forces the model to learn more robust feature representation by randomly discarding some neurons, and obtain the output , define the Dropout rate as p, then the operation of the Dropout layer is expressed as: ； S3305. The output of the Dropout layer will be input to the fully connected layer, which will further process the features through weighting and biasing, and finally obtain the probability of traffic accidents at each moment P. accident (t), the output layer is calculated as: ； Where W f represents the weight matrix of the fully connected layer, which is used to weight the hidden state of the input; b f Represents the bias term, which is used to adjust the output of the model.

5. The intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to claim 1 is characterized in that: The process of dynamically adjusting the warning threshold is as follows: ； Where θ(t) represents the dynamically adjusted warning threshold; θ fixed represents the set initial fixed threshold; α represents the weighting coefficient of historical accident data; represents the change in the historical accident rate; β represents the traffic density weighting coefficient; D traffic (t) represents the current traffic density; γ represents the weighting coefficient of weather factors; W weather (t) indicates the current weather conditions.

6. The intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to claim 1 is characterized in that: The rules for classifying warning levels are as follows: Red warning, i.e., serious accident risk: When the predicted probability exceeds 90% of the dynamic threshold θ(t), a red warning is triggered, indicating a very high accident risk and requiring immediate emergency response measures: P accident (t)>0.9×θ(t); Orange warning, i.e. higher accident risk: When the predicted probability is between 80% and 90%, an orange warning is triggered, indicating a higher accident risk: 0.8×θ(t) ≤P accident (t)≤0.9×θ(t); Yellow warning, i.e. medium accident risk: When the predicted probability is between 60% and 80%, a yellow warning is triggered, prompting the driver to drive cautiously and there is a certain accident risk: 0.6×θ(t) ≤P accident (t)<0.8×θ(t); Green warning, which means low accident risk: When the predicted probability is less than 60%, it means the probability of an accident is low, indicating smooth traffic: P accident (t)<0.6×θ(t).

7. An intelligent traffic accident prediction and emergency response system based on vehicle-mounted data, which is used to execute the intelligent traffic accident prediction and emergency response method based on vehicle-mounted data according to any one of claims 1 to 6, characterized in that: include: The vehicle data acquisition module is used to collect vehicle driving data through vehicle-mounted equipment and obtain environmental data through vehicle wireless communication technology V2X communication; On-board data processing module, used to improve data quality by using data pre-processing technology; Intelligent traffic accident prediction module, which is used to build an intelligent prediction model, conduct accident risk assessment on current traffic conditions, and calculate the probability of accidents occurring in the future; Intelligent early warning and risk warning module, used to generate early warning information based on the risk assessment results of the intelligent prediction model; The emergency response dispatch module is used to automatically or manually initiate emergency measures according to the situation after the warning is triggered; The traffic accident data analysis and optimization module is used to continuously optimize the prediction accuracy by collecting and analyzing data before the accident occurs, and use reinforcement learning technology to adjust the accident response strategy and dynamically adjust the traffic flow control strategy.

Images