Development of Vehicle State Estimation Method for Dedicated Sensor-Less Semi-active Suspension Using AI Technology

Abstract

This paper presents a sensor-less vehicle state estimation method using a neural network for semi-active suspensions. This method surpasses conventional mathematical models in performance and reduces calibration effort. The developed system, logic, and learning method are designed to address AI-specific challenges such as increased processing load and learning techniques, and their performance is validated through simulations and real-world tests. The results show that this system performs on par with those using dedicated sensors.

1 Introduction

Recently, various sensor configurations for semi-active suspension systems have been mass-produced. Among these, sensor-less systems, which use wheel speed data instead of dedicated sensors to reduce costs, remain relatively rare [1,2,3,4]. This study applies neural networks (NN) to a sensor-less system that omits conventional dedicated sensors, proposing a solution that combines high performance with low cost [5, 6].

2 Sensor-Less AI Semi-active Suspension System

2.1 Overview of Sensor-Less AI Semi-active Suspension

Figure 1 illustrates the sensor-less AI semi-active suspension logic configuration. The system estimates vehicle states using wheel speeds without sensors like G-sensors. Recurrent neural networks (RNN) replace formula-based logic, using CAN signals such as wheel speeds that vary with road surface inputs. The system outputs the vehicle body's vertical speed and the suspension's stroke speed, eliminating the need for dedicated sensors and enabling a cost-effective semi-active suspension system that enhances ride comfort.

Fig. 1.

AI semi-active suspension logic configuration

2.2 Learning Data

This system is designed to control frequencies related to primary and secondary rides. Consequently, the learning data focuses on vehicle vibrations within the 0.8 to 5 Hz range. Figure 2 illustrates the frequency distribution of the training data.

Fig. 2.

An example of the frequency distribution of the training data

3 Features of AI Semi-active Suspension

3.1 Using Synthetic Data to Enhance Ride Comfort

A notable advantage of AI is the ability to use synthetic data for learning. Synthetic data is artificially generated rather than naturally occurring. Figure 3 depicts the NN learning flow.

Fig. 3.

NN learning flow with synthetic data

Using a test vehicle, we measure CAN data such as wheel speed and the corresponding vehicle behavior. The measured vehicle behavior undergoes frequency axis integration, allowing the calculation of body speed and piston speed with optimal frequency characteristics for artificial control. Training the NN with this CAN data and the artificially generated vehicle behavior results in an NN with characteristics that are difficult to achieve with traditional formulas. The integral characteristic is shown in Figure 4.

Fig. 4.

Integral characteristics

The integral filter characteristic used in formulas reduces low-frequency gain to minimize the impact of gradients and sensor tilt. However, this causes the control band's phase to shift from the ideal characteristics. In contrast, the frequency axis integral characteristic prevents phase shift in the control band while reducing low-frequency gain. This improvement reduces the impact of gradients and enhances ride comfort near the sprung resonance frequency.

3.2 Enhanced Robustness with Relative Wheel Speed

AI robustness is a significant challenge, requiring accurate state estimation even under unlearned conditions. By setting the input signal's wheel speed as a relative value to other wheels rather than an absolute value, it becomes easier to detect wheel speed variations caused by road surface inputs. This approach improves robustness across various unlearned road surfaces and vehicle speeds.

4 Result

4.1 NN Estimation Results

Figure 5 displays the measured and estimated vertical body velocity. The NN provides high estimation accuracy over a wide range by learning the suspension's characteristics.

Fig. 5.

Measured and estimated vertical body velocity of uneven road

Figure 6 shows the transfer function of the learning data and the NN output. Due to the small amplitude of the NN output, a gain error occurs in the control band. However, the phase at the sprung resonance frequency matches the learning data, exhibiting the intended phase characteristics.

Fig. 6.

Transfer function of the learning data and the NN output

5 Ride Comfort Evaluation Results

Figure 7 shows the floor displacement, pitch angle, and sprung acceleration PSD when driving on a swell road with the AI system without sensors and a non-AI system with sensors. The AI system's displacement and sprung acceleration PSD match those of the sensor-equipped system, but the pitch angle is smaller, enhancing ride comfort.

Fig. 7.

Floor displacement, pitch angle and floor acceleration PSD when driving a swell road with AI system without sensor and non-AI system with sensor

Figure 8 illustrates the floor displacement, pitch angle, roll angle, and sprung acceleration PSD when driving on an uneven road with the AI system without sensors and a non-AI system with sensors. The AI system's displacement, roll angle, and sprung acceleration PSD are similar to those of the sensor-equipped system, but the pitch angle is smaller, enhancing ride comfort.

Fig. 8.

Floor displacement, pitch angle and floor acceleration PSD when driving a uneven road with AI system without sensor and non-AI system with sensor

6 Conclusion

We have developed a dedicated sensor-less semi-active suspension system using AI and confirmed its performance surpasses that of sensor-equipped systems.