Bus Driving Behavior Evaluation System Based on Big Data Machine Learning

1 Background

Driver behavior performance has a significant impact on safety and energy consumption performance. A large number of traffic accident statistics show that about 90% of accidents are caused by human error, resulting in a large number of casualties and economic losses each year. In the case of the same vehicle model, driving technology is the most critical factor in automotive energy consumption. The Standardization Administration of China has tested and verified that drivers at different driving levels have a fuel consumption difference of up to 8%-15% under the same conditions of use.
As a result, improving and standardizing driving behaviors can help reduce traffic accidents and energy consumption for businesses, which is of great social and economic value.

2 Content of Driving Behavior Evaluation System Establishment

Currently, most driving evaluations are driving behavior event statistics or manual weight setting, and evaluation on a driver is not objective. Yutong expects to build an objective and intelligent evaluation system of bus driving behavior based on big data analysis and machine self-learning algorithm model. Specifically, a driving behavior evaluation algorithm system process is established, including a whole process of data quality correction, feature engineering, feature verification, model building, and model optimization. Through the large data machine self-learning and gradient descending algorithms, the relativity and score weights of driving behavior indexes and customer actual operation conditions are continuously calculated and optimized, model training is perfected, and the subjective evaluation is avoided, so as to obtain objective driving score results and improvement suggestions.

2.1 Building of Large Data Driving Evaluation Model

To evaluate the driving behavior, the overall process of the evaluation model includes six processes, as shown in the figure below:

Figure Establishment of Big-data Driving Behavior Evaluation Model

  1. Set evaluation objectives: Provide results of evaluation objectives to customers for evaluation and ranking of relevant standards, such as safety evaluation and energy consumption evaluation.
  2. Dimension design: Each evaluation target is decomposed, and a behavior dimension that has the highest correlation with the target is sorted out. For example, dimensions that affect safety include violation of rules, unsmooth operation, and distraction of drivers. Dimensions related to fuel consumption include unsmooth operation and improper use of air conditioner.
  3. Element analysis: Analyze the related elements of each dimension. For example, the violation that endangers driving safety includes running red lights, over-speed, lane changing at the solid line, and driving beyond a specified lane. Unsmooth manipulation includes rapid acceleration, rapid deceleration, and sharp turning.
  4. Data source analysis: Determine data sources required by each element. For example, to obtain the fast acceleration data, the GPS speed data, the TBOX wheel speed data, or the acceleration data of the three-axis acceleration sensor is required.
  5. Indicator Validation: Include indicator definition and indicator validation. The behavior factors that have been refined in element analysis are not statistical indicators for direct calculation so that the elements need to be processed accordingly. For example, indicators such as "number of running red lights in the last three months" and "number of acceleration in a hundred kilometers" can be defined for the comprehensive evaluation of drivers. In addition, the validity of indicators needs to be verified based on the statistical test method, and invalid indicators need to be removed.
  6. Model building: To obtain the final score, we need to determine the relative weights between indicators and between targets.

2.1.1 Data Prepocessing

Data collected from different data sources is different in accuracy. In order to improve the utilization of data, the system adopts the corresponding method for accuracy improvement. For high frequency GPS and sensor data, an improved particle filtering algorithm is proposed, as follows:

In order to improve the tracking capability of the filter and robustness of change of process parameters, a particle filter algorithm of multi-data source fusion is used. Probability-based multi-source data fusion can effectively filter out random fluctuations from a single source or continue to be over/under-estimated. As shown in the figure below, after passing through the filter, the noise of the sensor data is removed effectively, and the trend data becomes clearer.

2.1.2 Data Indicator Validation

Indicator verification refers to verifying whether indicators are associated with our evaluation objectives. More specifically, it is to check whether the difference of indicators among users with different evaluation scores is significant statistically.
The significance test is one of the "Statistical hypothesis testing". The significance test mainly verifies whether the selected data indicators affect the evaluation results.
Through the significance test, we can determine the relationship between indicators and objectives one by one and lay the foundation for the final model.

2.1.3 Model Building

Model building refers to determining the relative weights between indicators and between targets in order to obtain the final score. In the model-based approach, objectives such as "fuel consumption" and " safety" have clear evaluation criteria so that the weight evaluation between indicators applies. The data flow for the model is shown below.

2.2 Establishment of Driving Behavior Application Platform

Based on the evaluation results of the large data driving evaluation model, we provide the corresponding product applications, including:

  • Overview of Enterprise Scores allows for viewing and analyzing scores of enterprises
  • The Details of Driver Scores allows for viewing the driver scores and improvement suggestions, and analyzing the influencing factors.
  • Summary of Driving Behavior allows for viewing details of driving behavior and analyzing scoring basis.

2.3 System Performance

Currently, driving behavior evaluation systems have been applied in vehicles by many transport enterprises in Hangzhou, Macao, Quanzhou, Ezhou and other areas in China. After seven months' trial running, the average fuel consumption of these sample vehicles decreases by 12% and the number of safety events decreases by 11% as the score level increases. If the annual mileage is 50,000 kilometers per vehicle, improving the driving behavior as recommended will save energy by 2L per hundred kilometers. Theoretically, it can save about 6,800 yuan (€850 or $970) per vehicle every year.

3 Evolution of Subsequent Models

  1. With popularization of intelligent devices such as ADAS, more accurate and timely data are subsequently used for driving safety evaluation.
  2. Integrate regional factors. For example, the same driving behavior shows different levels of safety and energy consumption in different regions in China. The purpose is to further improve the rationality of our driving behavior evaluation solution.
  3. Integrate scenario factors that are further refined. For example, the same driving behavior shows different levels of safety and energy consumption at different climatic conditions, different time, and different places. The purpose is to further improve rationality of our driving behavior evaluation solutions.
Comments (login to comment)
Receive our newsletter
Subscribe to our monthly newsletter and stay up to date! Receive the latest news, hot topics, seminar invitations and much more!
By submitting your emailaddress, you agree to our privacy policy.
Create your account now!
  • comment on articles
  • download presentations
  • stay up to date
Create your account!
Want to know more?

Want to know more about Busworld Academy, an article, an expert, a seminar, ...?

Ask your question