Method for predicting rollover risk of truck on sand-accumulated road surface based on machine learning
The machine learning-based method employing a Transformer-LSTM model addresses the challenge of predicting truck side flipping on sandy roads by simulating vehicle dynamics and identifying risk thresholds, effectively reducing accident rates by enhancing prediction accuracy and robustness.
Patent Information
- Application Number
- CN202411765387.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-04
- Publication Date
- 2025-03-21
AI Technical Summary
Current risk prediction methods fail to effectively address the increased likelihood of truck side flipping on sandy roads due to the unique combination of factors such as circular radius, slope, and friction coefficient changes, which pose significant challenges for vehicle stability and safety.
A machine learning-based method using a Transformer-LSTM model to predict the risk of truck side flipping on sandy roads by simulating vehicle dynamics and processing normalized data sets to identify risk thresholds and classify hazards, thereby enhancing prediction accuracy and robustness.
The method significantly reduces the occurrence of accidents by accurately modeling and predicting truck side flipping risks on sandy roads, leveraging the strengths of Transformer and LSTM models to improve prediction accuracy and robustness.
Smart Images
Figure BDA0005168802730000051 
Figure BDA0005168802730000052
Abstract
Description
Technical field:
[0001] The present invention belongs to the technical field of traffic safety prediction, and in particular relates to a method for predicting the risk of truck rollover on a sandy road surface based on machine learning. Background technology:
[0002] Highway bends are frequent sections of traffic accidents, with a high probability of occurrence and high severity. The risk of vehicle rollover increases with the increase of mass and vehicle size. Compared with buses, trucks have a much higher probability of accidents due to their large mass, high center of mass, and large inertia. In addition, truck traffic accidents cause great social harm and high economic losses. At present, deep learning and machine learning algorithms are being used in the field of road safety to predict safety hazards in extreme environments, but current research on extreme environments mainly focuses on severe weather such as rain, snow, and crosswinds. However, to date, many roads have crossed or are adjacent to China's eight major deserts and four major sandy areas. These sand-crossing roads have not only changed the previous traffic conditions around the edge of the desert, shortened the driving time between the two deserts, and greatly improved China's road traffic network system. At the same time, the current risk prediction research mainly focuses on the impact of the entire road environment on driving safety. In fact, different circular curve radii and transverse and longitudinal slope designs have different effects on driving safety. The driving safety of vehicles at curves is affected by the coupling of multiple factors such as circular curve radius, superelevation rate, slope, deflection angle, vehicle speed, friction coefficient, etc. At the same time, the accumulation of sand on the road causes a sudden change in the road friction coefficient, thereby affecting the driving stability of the vehicle. Therefore, it is urgent to design a method for predicting the risk of truck rollover on sandy roads. Summary of the invention:
[0003] In view of the current lack of heavy truck rollover risk prediction under sandy road conditions, the present invention designs a method for predicting truck rollover risk on sandy roads based on machine learning to achieve the purpose of reducing the incidence of accidents on sandy roads.
[0004] A method for predicting the risk of truck rollover on a sandy road based on machine learning, the method comprising the following steps:
[0005] Constructing original data sets: constructing a "people-vehicle-road" closed-loop simulation model, selecting vehicle operating parameters as inputs of the closed-loop simulation model, outputting lateral acceleration and lateral load transfer rate, taking the vehicle operating parameters as the first original data set, and taking the vehicle operating parameters, lateral acceleration and lateral load transfer rate as the second original data set;
[0006] Preprocessing the original data set: normalizing the first and second original data sets respectively, encoding and dimensionally expanding the normalized first and second original data sets respectively to complete preprocessing; dividing the preprocessed first and second original data into a first training set, a second training set, a first test set, and a second test set according to a ratio;
[0007] Constructing a risk regression model: constructing a Transformer-LSTM regression model, training the Transformer-LSTM regression model with the first training set to obtain a risk regression model, and verifying the risk regression model with the first test set;
[0008] Constructing a risk prediction model: inputting the first and second original data sets into the clustering analysis model, performing clustering after determining the optimal number of clusters, and outputting the clustering results as risk classification, and determining the thresholds of various types of risks; constructing a Transformer-LSTM classification model, training the Transformer-LSTM classification model with the second training set to obtain a risk prediction model, and verifying the risk prediction model with the second test set.
[0009] Preferably, the Transformer-LSTM regression model includes a first Transformer encoder, an LSTM layer, a global average pooling layer, a fully connected layer, a dropout layer and an output layer.
[0010] Preferably, the first Transformer encoder includes layer normalization, a multi-head attention mechanism and a feedforward neural network.
[0011] Preferably, the Transformer-LSTM classification model includes a second Transformer encoder, an LSTM layer and a fully connected layer.
[0012] Preferably, the second Transformer encoder includes layer normalization and a multi-head attention mechanism.
[0013] Preferably, the cluster analysis model is a K-means model, and the elbow rule is used to determine the optimal number of clusters.
[0014] Preferably, a "man-vehicle-road" closed-loop simulation model is constructed by trucksim simulation software.
[0015] Preferably, the vehicle operating parameters are vehicle speed, friction coefficient, load, circular curve radius, superelevation, and slope.
[0016] Preferably, the original data is normalized by MinMaxScaler, and the normalized original data is encoded and expanded by LabelEncoder.
[0017] Preferably, the raw data is normalized using the standard deviation method via MinMaxScaler.
[0018] In view of the characteristics of sandy roads, the present invention designs a method for predicting the risk of truck rollover on sandy roads based on machine learning, which supplements the factor of road sand accumulation in the existing risk prediction field, enriches the method of predicting the rollover of heavy trucks in harsh environments, and can effectively reduce the accident rate. The present invention uses the improved fusion of Transformer and long short-term memory neural network (LSTM) as the main risk prediction algorithm. The vehicle dynamics data of heavy trucks under different road sand accumulation conditions are obtained through numerical simulation as source data, and risk classification and regression analysis are carried out based on these source data, and finally risk prediction is performed according to the risk classification and regression results. The Transformer-LSTM model combines the advantages of the Transformer and LSTM models, and dropout is applied in the construction process of the model to avoid overfitting. The model has been greatly improved in terms of accuracy and robustness. Specific implementation method:
[0019] In order to make the technical solution of the present invention easier to understand, the technical solution of the present invention is now clearly and completely described in the form of specific embodiments.
[0020] Construct the original data set: Use stucksim simulation software to construct a "people-vehicle-road" closed-loop simulation model, select vehicle speed (40km / h, 50km / h, 60km / h, 70km / h, 80km / h, 90km / h), friction coefficient (0.675, 0.76, 0.785, 0.95, 1.09, 1.19, 1.28, 1.38), load (5t, 15t, 25t, 30t, 44t, 45t), circular curve radius (60m, 100m, 150m, 200m, 2 The vehicle operating parameters such as height (15m, 300m, 400m), superelevation (1%, 2%, 3%, -1%, -2%, -3%), and slope (0%, 1%, 2%, 3%, 4%) are used as inputs of the closed-loop simulation model for multi-factor experiments, and lateral acceleration and lateral load transfer rate are output. The vehicle dynamics data of heavy trucks under different road sand conditions are obtained by numerical simulation as source data, and the vehicle operating parameters are used as the first original data set, and the vehicle operating parameters, lateral acceleration and lateral load transfer rate are used as the second original data set;
[0021] Preprocessing the original data set: The first and second original data sets are normalized by MinMaxScaler using the standard deviation method. Normalization is a way to simplify calculations, that is, to transform dimensional expressions into dimensionless expressions and become scalars. The normalized first and second original data sets are encoded and dimensionally expanded by LabelEncoder to allow the data to complete preprocessing so that they are suitable for the subsequent Transformer encoder and LSTM layer; the preprocessed first and second original data are divided into the first training set, the second training set, the first test set, and the second test set in a ratio of 8:2; among them, the vehicle operating parameters are used as feature data, and the lateral acceleration and lateral load transfer rate are used as target data. By learning the training set, the model can find the relationship between the input features and the target output, thereby adjusting its internal parameters to minimize the prediction error. During the training process, the model will continuously update its parameters and use the data in the training set to improve the accuracy of the prediction.
[0022] Constructing a risk regression model: constructing a Transformer-LSTM regression model, training the Transformer-LSTM regression model with the first training set to obtain a risk regression model, and verifying the risk regression model with the first test set; the Transformer-LSTM regression model includes a first Transformer encoder, an LSTM layer, a global average pooling layer, a fully connected layer, a dropout layer, and an output layer. The Transformer-LSTM regression model passes the output of the encoder to the LSTM layer to capture sequence features, and converts it into a fixed-length vector through a global average pooling layer, and then passes through multiple fully connected layers and dropout processing to finally generate an output layer; wherein, the encoder includes layer normalization, a multi-head attention mechanism, and a feedforward neural network, and uses residual connections to enhance feature transfer.
[0023] Constructing a risk prediction model: inputting the first and second original data sets into the K-means clustering analysis model, clustering after determining the optimal number of clusters using the elbow rule, and outputting the clustering results as risk classification, determining the thresholds of various risks (as shown in Table 1) and various risk levels (as shown in Table 2), and taking the risk level as the target; constructing a Transformer-LSTM classification model, training the Transformer-LSTM classification model with the second training set to obtain a risk prediction model, and verifying the risk prediction model with the second test set; the Transformer-LSTM classification model includes a second Transformer encoder, an LSTM layer, and a fully connected layer. The encoder includes layer normalization and a multi-head attention mechanism for extracting sequence features, passing the output of the encoder to the LSTM layer for sequence processing, and finally performing feature extraction and classification output through the fully connected layer.
[0024] Table 1 Risk indicator thresholds
[0025]
[0026] Table 2 Risk levels
[0027]
[0028] This embodiment also discloses a method for retraining the model to increase the model prediction performance:
[0029] Deploy traffic cameras around the road to collect data on circular curve radius, superelevation, friction coefficient, load, speed, and slope of vehicles, which are recorded as new parameters;
[0030] The collected new vehicle operation parameters are preprocessed, and the preprocessed parameters are input into the risk regression model to predict new data, thereby obtaining predicted new data (lateral acceleration, lateral load transfer rate), and a new data set is formed by the new parameters, new lateral acceleration and lateral load transfer rate;
[0031] The new data set is merged with the first and second preprocessed original data sets to form a new original data set, the new original data set is input into the k-means clustering analysis model to output a new risk classification result, the new risk classification result and the new original data set are input into the risk prediction model for retraining, and the new data set is input into the retrained risk prediction model for risk prediction. In real life, the risk prediction model can be trained using this method by continuously collecting new data to improve the prediction accuracy of the method disclosed in the present invention.
[0032] It should be noted that the embodiments described herein are only partial embodiments of the present invention, rather than all implementations of the present invention, and the embodiments are only exemplary, and their role is only to provide a more intuitive and clear way to understand the content of the present invention, rather than to limit the technical solutions described in the present invention. Without departing from the concept of the present invention, all other implementation methods that can be thought of by ordinary technicians in this field without creative work, and other simple replacements and various changes to the technical solutions of the present invention, all belong to the protection scope of the present invention.
Claims
1. A method for predicting the risk of truck rollover on sandy roads based on machine learning, characterized in that: The method comprises the following steps: Constructing original data sets: constructing a "person-vehicle-road" closed-loop simulation model, selecting vehicle operating parameters as inputs of the closed-loop simulation model, outputting lateral acceleration and lateral load transfer rate, taking the vehicle operating parameters as the first original data set, and taking the vehicle operating parameters, lateral acceleration and lateral load transfer rate as the second original data set; Preprocessing the original data set: normalizing the first and second original data sets respectively, encoding and dimensionally expanding the normalized first and second original data sets respectively to complete preprocessing; dividing the preprocessed first and second original data into a first training set, a second training set, a first test set, and a second test set according to a ratio; Constructing a risk regression model: constructing a Transformer-LSTM regression model, training the Transformer-LSTM regression model with the first training set to obtain a risk regression model, and verifying the risk regression model with the first test set; Constructing a risk prediction model: inputting the first and second original data sets into the clustering analysis model, performing clustering after determining the optimal number of clusters, and outputting the clustering results as risk classification, and determining the thresholds of various types of risks; constructing a Transformer-LSTM classification model, training the Transformer-LSTM classification model with the second training set to obtain a risk prediction model, and verifying the risk prediction model with the second test set.
2. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: The Transformer-LSTM regression model includes a first Transformer encoder, an LSTM layer, a global average pooling layer, a fully connected layer, a dropout layer and an output layer.
3. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 2, characterized in that: The first Transformer encoder contains layer normalization, multi-head attention mechanism and feedforward neural network.
4. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: The Transformer-LSTM classification model includes a second Transformer encoder, an LSTM layer, and a fully connected layer.
5. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 4, characterized in that: The second Transformer encoder contains layer normalization and multi-head attention mechanism.
6. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: The cluster analysis model is a K-means model, and the elbow rule is used to determine the optimal number of clusters.
7. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: A "people-vehicle-road" closed-loop simulation model is constructed using trucksim simulation software.
8. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: The vehicle operating parameters are vehicle speed, friction coefficient, load, circular curve radius, superelevation, and slope.
9. The method for predicting the risk of truck rollover on a sandy road based on machine learning as claimed in claim 1, characterized in that: The original data is normalized by MinMaxScaler, and the normalized original data is encoded and expanded by LabelEncoder.
10. The method for predicting the risk of truck rollover on sandy roads based on machine learning as claimed in claim 9, characterized in that: The raw data were normalized using the standard deviation method using MinMaxScaler.