Estimation of Hybrid Deep Learning-Based Fail-Safe Control for Four-Wheel Steering Systems ^†

Abstract

Functional safety in a four-wheel steering system is critical for smart vehicles, especially under steer-by-wire configurations where conventional methods fail to achieve 100% safety. In this study, we developed a hybrid AI-based fail-safe control framework that combines long short-term memory and convolution neural network classifiers for real-time fault detection, prediction, and safe state decision-making. The developed system was trained using steering sensor data, incorporating accurate response modeling and fault induction mechanisms. An LSTM model predicts deviations in steering behavior, while a random forest (RF) classifier identifies fault types and initiates safe state transitions in accordance with ISO 26262 functional safety guidelines. The system includes a Texas Instruments TMS320F28377D microcontroller and a rear steering actuator with a permanent magnet synchronous motor, utilizing controller area network communication for diagnostics and actuation. The system presents the potential of machine learning techniques to enhance the fail-safe capabilities of next-generation steer-by-wire systems.

1. Introduction

With the increasing integration of intelligent systems in vehicles, ensuring the reliable and fail-safe control of critical components such as steering motors has become an essential requirement, especially under the ISO 26262 automotive functional safety standard [1]. Four-wheel steering (4WS) systems enhance vehicle stability and maneuverability, but their reliance on precise motor control makes them vulnerable to sensor failures, actuator degradation, and external disturbances. A failure in motor behavior severely compromises safety, especially in autonomous or high-assist driving scenarios [2,3,4].

To address these challenges, we developed a novel deep learning-based motor control system for 4WS systems that satisfies ISO 26262 safety goals. The developed system leverages a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) architecture to manage motor torque and angle control in real time, even in the presence of sensor faults or uncertainty [5]. The CNN components efficiently extract spatial features from raw sensor inputs such as speed, acceleration, and current signals, while LSTM captures temporal dependencies critical to dynamic motor behavior and vehicle response [6].

Unlike traditional proportional–integral–derivative or model-based control strategies, the hybrid CNN–LSTM controller enables non-linear, time-varying dynamics, enabling intelligent decision-making and real-time adaptation under fault conditions. To enhance safety compliance, the control framework incorporates a sequential probability ratio test (SPRT) module for fault detection. This statistical layer monitors the residuals between expected and observed motor responses and ensures timely transitions to fail-operational modes in the event of anomalies [7,8,9,10].

The developed system improves tracking accuracy and system stability and embeds redundancy and diagnostic mechanisms aligned with the ISO 26262 Automotive Safety Integrity Level requirements. Testing results, including simulation and hardware-in-loop validations, demonstrate the system’s ability to maintain safe and effective motor control during both normal operation and fault scenarios [11].

By introducing an ISO 26262-oriented, AI-enhanced motor control strategy for 4WS systems, the learning capacity of hybrid CNN–LSTM models is integrated with real-time fault detection, offering both performance and safety in a unified framework (Figure 1).

2. Methodology for Motor Speed Control

The developed hybrid CNN-LSTM-based motor control system for four-wheel steering is designed to operate with high reliability under ISO 26262 functional safety requirements. The design includes system initialization, activating all necessary subsystems as shown in Figure 2, such as the communication interface, neural network engine, and fault detection routines [12]. The control system receives critical input data via the controller area network bus, which includes motor feedback signals such as angular speed, torque commands, and other vehicle dynamics information.

Once the data are acquired, the system identifies the real-time speed of the motor by analyzing these inputs. These preprocessed data are then transmitted through a hybrid deep learning model composed of CNN and LSTM layers. The CNN component extracts spatial features from multi-sensor input patterns, while the LSTM component captures the temporal dependencies and dynamics of the motor behavior. Collectively, they are used to predict the expected motor response under current operating conditions. Simultaneously, the system performs a real-time comparison between the predicted motor speed and the measured signal. This functions as a fault detection mechanism to determine whether the motor is operating within safe bounds. If a potential fault is detected, the system decodes the fail-safe logic and transitions to a fallback mode. In this mode, the faulty motor speed input is replaced with the predicted value of the CNN–LSTM, and the motor speed is adjusted accordingly to ensure continued but safe operation [13,14,15,16]. Finally, the updated motor control signals are sent to the actuator, maintaining the correct motor behavior while preserving safety and stability. This closed-loop framework enables intelligent fault detection, robust motor control, and fail-safe response, fulfilling the real-time safety objectives of ISO 26262 for safety-critical vehicle systems.

3. Model Analysis

To enable intelligent fail-safe control in four-wheel steering motor systems, a hybrid deep learning architecture combining CNN and LSTM is employed. A mathematical model was analyzed for the design and implementation of the hybrid estimator, along with the sensor integration and performance evaluation metrics in the experiment (Figure 3).

3.1. Data Acquisition and Preprocessing

The system collects real-time sensor data, including motor angular velocity ω(t), motor torque τ(t), vehicle speed v(t), and acceleration a(t), from onboard sensors such as encoders and inertial measurement units. These signals are normalized using the minimum–maximum scale. This ensures consistent input scaling for the neural network and stabilizes training.

$$x_{norm}=-{\displaystyle \frac{x-x_{min}}{x_{max}-x_{min}}}$$

(1)

where $x_{norm}$ is the normalized scaling, x is the data input, $x_{min}$ is the minimum data value, and $x_{max}$ is the maximum data point.

3.2. CNN–LSTM Hybrid Architecture

The hybrid model is composed of 1D convolutional layers followed by LSTM layers. CNN extracts spatial features from the input signal window X, where n is the time window and m is the number of features. The convolution operation is defined as follows.

$$y_i=\sum _{j=0}^{\omega -1}{x}_{i+j}·w_j+b$$

(2)

Here, $w_j$ is the filter weight for the kth kernel, $b$ is the bias term, $x_{i+j}$ is the input segment, and $y_i$ is the output feature map. These extracted features are then passed to the LSTM unit, which models temporal dependencies. The LSTM cell operations are given by the following Equations (3)–(8).

$$\mathrm{Forget} \mathrm{gate} \left(f_t\right)=\sigma \left(w_f·\left[h_{t-1,}{x}_{,}\right]+b_f\right)$$

(3)

$$\mathrm{Input} \mathrm{gate} \left(i_t\right)=\sigma \left(w_i·\left[h_{t-1,}{x}_{,}\right]+b_i\right)$$

(4)

$$\mathrm{Candidate} \mathrm{cell} \mathrm{gate} \left({\check{c}}_t\right)=tanh\left(w_c·\left[h_{t-1,}{x}_{,}\right]+b_c\right)$$

(5)

$$\mathrm{Updated} \mathrm{cell} \mathrm{gate} \left(c_t\right)=f{c}_{t-1}+i_t·{\check{c}}_t$$

(6)

$$\mathrm{Outputgate} \left(o_t\right)=\sigma \left(w_o·\left[h_{t-1,}{x}_{,}\right]+b_o\right)$$

(7)

$$\mathrm{Hidden} \mathrm{state} \left(h_t\right)=o_t\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(c_t\right)$$

(8)

Here, $w_f$,$w_i$ ,$w_c, w_o, b_f, b_i, b_c, \mathrm{and} b_o$ are the weights and biases of the forget gate, input gate, candidate gate, and output gates, respectively. σ is the sigmoid activation function, $i_t$ is the in input of the t index, $c_{t-1}{ \mathrm{is} \mathrm{the} \mathrm{previous} \mathrm{cell} \mathrm{state}, h}_{t-1}$ is the previous hidden state and is the output of LSTM at a time step. ${\check{c}}_t$ and $c_t$ are the cell current and updated states.

The residual in the prediction is calculated as follows:

$$r\left(t\right)= \omega \left(t\right)- \widehat{\omega }\left(t\right)$$

(9)

where the predicted motor speed $\widehat{w}$(t) is compared to the actual measured speed $\widehat{\omega }$(t) to compute the residual. The model accuracy is assessed using the root mean squared error (RMSE) and mean absolute error (MAE). The plotted training curves demonstrate convergence and validation accuracy, validating model robustness. This residual is used in the sequential probability ratio test (SPRT) for fault detection.

$$\mathrm{MAE}={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\left|t_i-{\widehat{t}}_i\right|$$

(10)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{(t_i-{\widehat{t}}_i)}^2}$$

(11)

where N is the total data points; i represents the index of each data point; $t_i$ is the actual measured value; and ${\widehat{t}}_i$ is the predicted value.

4. Results

The performance of the developed hybrid deep learning-based fail-safe motor control strategy was evaluated using both real-time data and predicted outputs. The key focus of the analysis was to examine the ability of the LSTM and CNN–LSTM models to accurately follow the motor speed command and correct deviations under fault conditions.

Motor Speed Prediction Without Fail-Safe Implementation

To evaluate the effectiveness of the developed hybrid CNN–CNN-LSTM-based fail-safe system, its behavior was analyzed both without and with the implementation of the fault-tolerant logic. Figure 4 presents the motor performance without any fail-safe mechanism. As shown in the left subfigure, although the live speed initially follows the command, there is a noticeable deviation in the latter stages, especially after 2000 s. This indicates that the motor is unable to maintain the commanded speed due to disturbance or fault conditions. Although the LSTM-adjusted output stabilizes the signal, the system exhibits persistent deviations from the commanded input, indicating the limitations of relying solely on a single deep learning model without incorporating structured fault isolation mechanisms.

In contrast, Figure 5 demonstrates the effectiveness of the developed fail-safe framework. With the CNN–LSTM model integrated alongside real-time monitoring, the system maintains an excellent alignment with the commanded speed even after 2000 s. The CNN–LSTM-adjusted speed shows a minimal deviation from the target, outperforming both the live signal and the LSTM-only correction. The inclusion of convolutional layers enhances spatial feature extraction, allowing the system to detect subtle sensor or actuator faults earlier. Moreover, the statistical fault detection mechanism, SPRT, triggers corrective control seamlessly by replacing faulty inputs with estimated values. As a result, the motor speed tracks the reference command much more accurately and consistently, validating the robustness and reliability of the fail-safe logic.

The developed hybrid deep learning model not only improves real-time speed prediction but also provides a strong resilience under fault scenarios, ensuring compliance with ISO 26262 safety expectations for critical systems such as four-wheel steering. To evaluate the prediction accuracy of the developed motor control models, the RMSE and MAE were calculated for each configuration, as summarized in

Table 1. RMSE and MAE in motor speed prediction.

Model	RMSE	MAE
LSTM (no fail-safe)	0.95	1.45
CNN-LSTM (no fail-safe)	0.81	1.05
LSTM (with fail-safe)	0.45	0.54
CNN-LSTM (with fail-safe)	0.54	0.15

. The LSTM model without a fail-safe logic shows relatively higher errors, with an RMSE of 0.95 and an MAE of 1.45, indicating more frequent and larger deviations between predicted and actual motor speeds. The introduction of a CNN layer enhances the baseline performance, as demonstrated by the CNN–LSTM model without fail-safe logic, which achieves an RMSE of 0.81 and a MAE of 1.05, indicating an improved learning of system dynamics. Performance gains are observed when fail-safe mechanisms are incorporated. Specifically, the LSTM model with fail-safe logic reduces the RMSE to 0.45 and the MAE to 0.54, confirming its increased robustness under fault conditions. The optimal configuration is achieved with the CNN–LSTM model integrated with fail-safe logic, yielding the lowest error metrics across all tested models: an RMSE of 0.54 and an MAE of 0.15. These findings underscore the efficacy of combining hybrid deep learning architectures with fault-tolerant strategies to sustain control performance in the presence of potential system failures, thereby aligning with the safety objectives outlined in ISO 26262.

5. Conclusions

We developed a hybrid deep learning-based system for fail-safe motor speed control in four-wheel steering systems to enhance reliability under ISO 26262 functional safety standards. By combining CNN with LSTM architectures, the system effectively learned both spatial and temporal features from motor-related sensor data to estimate and adjust motor speed under normal and faulty conditions. The integration of a fail-safe mechanism enables the real-time detection and compensation of faults by switching to predicted values from the hybrid model. The experimental results confirmed that the CNN–LSTM model with fail-safe logic significantly outperformed the other configurations, achieving the lowest RMSE and MAE, thereby ensuring a robust control and improved safety performance. The system developed demonstrates the potential of AI-driven control strategies to enhance both performance and resilience in safety-critical automotive systems. The system needs to include implementation on embedded automotive hardware, real-world testing, and integration with end-to-end vehicle control stacks.