Steering-by-wire angle error correction system based on neural network compensation

Abstract

The invention relates to the technical field of vehicle control, and discloses a steering-by-wire angle error correction system based on neural network compensation, the system adopts a time sequence error prediction integrated network to obtain a vehicle state and predict a mean value and a variance of future angle errors, and the variance is used for quantifying the uncertainty of prediction. And a main control module performs prospective compensation according to the mean value and realizes risk adaptive feedback control according to the variance. A correction module actively injects a detection signal into the steering system when the uncertainty exceeds a preset threshold, and corrects the network online according to the difference between actual and theoretical responses. According to the method, by establishing an intelligent closed loop of uncertainty perception-adaptive control-active learning, rapid, accurate and robust correction of steering errors is achieved, and the adaptive capacity and safety of the system under complex and changeable working conditions are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle control, in particular to a steer-by-wire angle error correction system based on neural network compensation. BACKGROUND

[0002] Steer-by-wire technology, as one of the core technologies of future intelligent vehicles, cuts off the traditional mechanical connection between the steering wheel and the steering wheels, and uses sensors, controllers and execution motors to transmit and execute the steering intention of the driver. This architecture not only provides great freedom for the internal design of the vehicle, but also paves the way for the integration of advanced driving assistance systems and autonomous driving functions.

[0003] In any steer-by-wire system, ensuring that the actual steering angle of the front wheels can accurately and quickly track the instructed steering angle issued by the driver or the upper controller is the primary prerequisite for ensuring the vehicle's handling performance and driving safety. However, in real driving environments, due to the nonlinearity of the vehicle's own dynamic characteristics, the complexity of the interaction between the tires and the road surface, and sudden disturbances from the external environment, tracking errors between the instructed steering angle and the actual steering angle are inevitable.

[0004] In order to suppress such errors, traditional control methods, such as proportional-integral-derivative controllers, are widely used. This type of controller is simple in structure and easy to implement, but has inherent limitations when dealing with complex working conditions. They are essentially a kind of lag correction, i.e. they can only compensate after the error has occurred, which leads to a delay in the system's response, and poor performance in scenarios that require fast response.

[0005] In order to overcome the response delay, some advanced control strategies based on models, such as model predictive control, are introduced into the steer-by-wire system. These methods use the vehicle's dynamics model to predict the future state of the system, thereby achieving anticipatory compensation for errors. However, the performance of such methods is highly dependent on the accuracy of the model built. The dynamics model of the vehicle is an extremely complex time-varying system, and its parameters will change constantly with factors such as tire wear, load changes, road conditions, etc. A fixed model calibrated under specific working conditions will have a significant decrease in prediction accuracy when faced with changes in working conditions or unmodeled dynamics, thereby weakening or even losing control effectiveness.

[0006] In recent years, the development of artificial intelligence technology, especially neural networks, has provided a new way to deal with such complex nonlinear problems. Through offline training on a large amount of driving data, a neural network controller can learn and fit the complex dynamic characteristics of the vehicle, and has stronger nonlinear expression ability than traditional physical models. However, this offline training-based mode also faces fundamental challenges. It assumes that the training data has completely covered all possible driving scenarios, but the complexity and suddenness of the real world make this assumption almost impossible to achieve. When the vehicle encounters a completely new working condition that has never appeared in the training data (for example, suddenly driving from dry asphalt pavement to icy pavement), the generalization ability of the pre-trained neural network model will be severely tested, and the accuracy of its output cannot be guaranteed.

[0007] A deeper defect is that existing control systems, whether based on physical models or data-driven models, usually lack the ability to judge their own "cognitive boundaries". When they output a control instruction, they cannot give a "confidence" evaluation of the effectiveness of the instruction at the same time. The system cannot judge whether it is currently in a "comfort zone" where the model is highly matched and predictable, or in a "danger zone" where the model has already been mismatched and full of uncertainty. This "insensitivity" to risk and uncertainty makes the system unable to dynamically adjust its control strategy according to the risk level, and even less able to take proactive measures to correct when it detects model failure, thus leaving hidden dangers for driving safety.

[0008] Therefore, how to design a steering error correction system that can perceive the uncertainty of its own model in real time and actively adapt to unknown working conditions according to the uncertainty is a key technical problem that needs to be solved in the current steer-by-wire technology field. SUMMARY

[0009] In view of the deficiencies of the prior art, the present application provides a steer-by-wire angle error correction system based on neural network compensation, aiming to improve the problem of control performance degradation and insufficient safety caused by the lack of real-time perception and active adaptation ability of the prediction uncertainty of the system when facing model mismatch and unknown disturbances in the existing steer-by-wire error correction method.

[0010] To achieve the above purpose, the present application is realized by the following technical scheme: a steer-by-wire angle error correction system based on neural network compensation, which can predict the probability distribution of future errors and actively detect the environment according to the predicted uncertainty to correct its own model online, thereby realizing fast, accurate and robust correction of the steering angle error.

[0011] The first aspect of the present application provides a steer-by-wire angle error correction system based on neural network compensation, comprising: an acquisition module, a prediction module, an analysis module, a main control module and a correction module.

[0012] The acquisition module is configured to acquire a real-time state vector of the vehicle in each control cycle. The real-time state vector can include the command steering angle, the actual steering angle, the current angle error, the vehicle speed, the yaw rate, and the historical error sequence, etc., to provide a comprehensive data basis for subsequent prediction and control.

[0013] The prediction module is connected with the acquisition module, and the core of the prediction module is a time series error prediction integrated network. The network is composed of multiple independent time series neural networks (such as gated recurrent unit GRU or long short-term memory LSTM network) in parallel. The real-time state vector is input into the integrated network, and each independent time series neural network independently outputs a prediction sequence of the angle error in the future. By using the integrated structure, the system can predict the future from multiple perspectives, which is the basis for quantifying uncertainty.

[0014] The analysis module is connected with the prediction module. The analysis module performs real-time statistical analysis on the multiple independent prediction sequences from the integrated network. Specifically, for a future time t+τ, the analysis module calculates the mean and variance of all prediction values.

[0015] The prediction error mean μ e (t+τ), which is an expected estimate of the most likely value of the future error, is calculated as follows:

[0016]

[0017] where N is the number of time series neural networks, is the angle error prediction value of the i-th network for the future time t+τ.

[0018] The prediction error variance as a core quantitative indicator of prediction uncertainty. The greater the variance, the greater the divergence of the network predictions, the more uncertain the system's understanding of the current working conditions, and the higher the risk it faces. The calculation method is as follows:

[0019]

[0020] The main control module is connected with the analysis module. The main control module generates a main control torque by using the mean and variance output by the analysis module through a double-channel parallel method. In one embodiment, the module first calculates the main control torque based on the prediction error mean μ e(t+τ) calculates a forward compensation torque, which can apply compensation before the error actually occurs, thus overcoming the response delay of traditional control. At the same time, the module also generates a risk adaptive feedback torque through a feedback controller. The core parameters of the feedback controller, such as the proportional, integral, and derivative gains of the PID controller, are no longer fixed values, but are designed as functions of the prediction error variance , so that the controller can automatically become more stable and conservative when the uncertainty is high (high risk), and pursue higher performance when the uncertainty is low.

[0021] The correction module, connected with the analysis module and the prediction module, constitutes one of the core innovations of the present application, namely the active detection and self-correction mechanism. The function of this module is reflected in the following aspects: first, it continuously judges whether the prediction error variance exceeds a preset activation threshold. When the variance is too high, it means that the prediction model is no longer accurate in describing the current working condition (such as sudden change of road adhesion coefficient). Second, if the threshold is exceeded, the correction module is activated and generates a high-frequency, small-amplitude active detection signal (for example, a specific disturbance torque) that is imperceptible to the driver, and injects it into the steer-by-wire system. Finally, the module observes the actual response of the system to this known active detection signal, and compares it with the theoretical response predicted by the time series error prediction integrated network. The difference between the two, i.e. the model mismatch error, directly reflects the degree of inaccuracy of the current prediction model. The correction module will fine-tune the internal weights of each time series neural network in the integrated network in real time and online according to this mismatch error through optimization algorithms such as gradient descent.

[0022] In one embodiment, the weight update rule can be represented as:

[0023]

[0024] where W i (t) is the weight of the i-th network at time t, η is the learning rate, L is the loss function based on the model mismatch error ε m,i (t), is the gradient of the weight.

[0025] Through this mechanism, the system no longer passively accepts environmental information, but actively "asks" the environment when it feels "confused" and corrects its self-awareness according to the "echo". This closed-loop design of active detection and online model correction enables the system to quickly adapt to new working conditions that have not been seen in the training data, fundamentally improving the robustness and adaptability of the system.

[0026] Finally, the main control torque generated by the main control module is superimposed with the active probing signal (generated by the correction module in the active state) to form a total torque output to the steering actuator, completing a control cycle.

[0027] The application further provides a storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the method.

[0028] The application provides a neural network compensation-based steer-by-wire angle error correction system.

[0029] 1. The application deploys a time series error prediction integrated network composed of multiple neural networks instead of relying on a single prediction model, and can obtain multiple independent predictions about future errors in parallel. Statistical analysis of these prediction results, especially the calculation of their variances, provides a key and quantifiable uncertainty index for the system. This enables the system to judge its prediction reliability for the first time, providing a solid and reliable decision basis for subsequent adaptive control and active learning.

[0030] 2. The application uses the mean of the prediction errors output by the integrated network, which represents the most likely estimate of the future error evolution trajectory. The feedforward compensation control channel based on this mean can actively cancel before the error is fully formed. This forward-looking correction method effectively overcomes the inherent delay of traditional feedback control, which must wait for the error to appear before responding, thereby achieving more rapid and accurate angle tracking, especially when the driver performs rapid steering operations.

[0031] 3. The application uses the quantified prediction uncertainty as a risk indicator and dynamically adjusts the key gain of the feedback controller. When the system senses an increased risk, it automatically switches to a more robust control mode, such as increasing the derivative gain to enhance system damping. This mechanism effectively avoids system oscillation or instability caused by excessive control behavior when the model is mismatched, ensuring driving safety.

[0032] 4. When the system senses that its prediction uncertainty exceeds a preset threshold, it no longer passively tolerates performance degradation, but actively injects a small probing signal into the physical world to obtain key dynamic feedback information. And update its internal prediction model in real time based on this information, so it can quickly adapt to model mismatch caused by sudden changes in road adhesion coefficient or changes in vehicle parameters, demonstrating excellent environmental adaptability.

[0033] 5、The application continuously and small-step corrects the prediction model, so that it closely tracks the current most real physical characteristics of the vehicle. This continuous self-optimization capability can automatically compensate for the effects of slow time-varying factors such as tire wear, suspension parameter changes, etc. on the vehicle dynamics model, avoiding the problem of gradual decline in performance over time caused by the inability to update the traditional fixed model, ensuring that the vehicle can maintain high-quality handling experience throughout its life cycle. BRIEF DESCRIPTION OF DRAWINGS

[0034] Fig. 1 It is an overall architecture diagram of the error correction system of the application.

[0035] Fig. 2 It is a flowchart of the error correction method of the application.

[0036] Fig. 3 It is an error distribution generation and double-channel control schematic diagram of the application. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the specification of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0038] Please refer to the drawings in the specification of the application Figs. 1 to 3 The embodiment of the application provides a steer-by-wire steering angle error correction system based on neural network compensation, which is deployed in a steer-by-wire vehicle platform. The hardware architecture of the platform provides a physical basis for the implementation of the method of the application.

[0039] The hardware architecture of the steer-by-wire system can specifically include a steering wheel assembly, a main controller, a steering execution assembly, and a vehicle sensor network.

[0040] The steering wheel assembly, as the input end of the driver's instruction, includes a steering wheel angle sensor and a steering wheel torque sensor for detecting the driver's intention. These sensors convert the driver's operation into digital signals and are the source of system instructions.

[0041] The main controller, as the operation core of the algorithm of the application, is a high-performance electronic control unit (ECU). It internally integrates a microprocessor, a random access memory (RAM), a non-volatile memory, and various communication interfaces. All the calculation steps involved in the application, including the inference and online update of the neural network, are performed in the main controller. The non-volatile memory is used to store the algorithm program and the initial model parameters of the timing error prediction integrated network.

[0042] The steering execution assembly is the final actuator to realize the steering of the wheels. It includes a high-power steering execution motor, a set of reduction mechanism to increase the torque, a rack and pinion mechanism to convert the rotary motion into linear motion, and a wheel angle sensor to accurately measure the actual wheel angle. The final control torque signal calculated by the main controller controls the steering execution motor through the drive circuit, thereby driving the wheel deflection.

[0043] The on-board sensor network provides the necessary vehicle dynamic information for the invention. The main controller obtains key information such as the longitudinal speed, yaw rate, lateral acceleration, etc. of the vehicle in real time from other control units of the vehicle (such as the vehicle body stability system ESP, engine control unit, etc.) through the controller area network (CAN) bus. These information together with the feedback signals from the steering wheel assembly and the steering execution assembly form the input of the algorithm of the invention.

[0044] A stable and reliable power supply system provides power for all the above-mentioned electronic components, ensuring the normal operation of the entire system.

[0045] Those skilled in the art can understand that the above-mentioned hardware architecture is only illustrative. The method of the invention is not limited to a specific hardware configuration, and any hardware platform that can provide similar functions and meet the algorithm operation requirements can be used as the implementation environment of the invention.

[0046] In a specific embodiment, the specific process of a steer-by-wire angle error correction method based on neural network compensation is realized by a program deployed in the main controller, and the method is executed in a fixed control cycle.

[0047] The first step is to collect real-time state information and construct a state vector. At the beginning of each control cycle t, the main controller collects a series of raw data through sensors and the CAN bus. After processing, these data are organized into a multi-dimensional real-time state vector X(t), which fully describes the dynamic characteristics of the current steering system.

[0048] In this embodiment, the state vector X(t) is defined as:

[0049] X(t) = [δ cmd (t), δ act (t), e δ (t), v x (t), ω z (t), T m (t-Δt), e δ (t-kΔt), …, e δ (t-Δt)] T

[0050] where δ cmd (t) is the driver's command steering angle acquired by the steering wheel angle sensor; δ act (t) is the actual front wheel steering angle fed back by the wheel angle sensor; e δ (t) is the current angle tracking error, i.e. e δ (t) = δ cmd (t) - δ act (t); v x (t) is the longitudinal driving speed of the vehicle; ω z (t) is the yaw rate of the vehicle; T m (t - Δt) is the actual torque outputted by the master controller to the steering motor in the last control period; {e δ (t - kΔt), …, e δ (t - Δt)} is a history error sequence containing the past k time steps, which is used to provide dynamic information of error evolution for the prediction network, where Δt is the length of the control period.

[0051] Secondly, the future error distribution prediction based on the integrated time series network. The core of this step is an integrated time series error prediction network (ETEP) deployed in the master controller, which receives the state vector X(t) as input and predicts the angle error distribution in the future.

[0052] The ETEP is composed of NN gated recurrent unit (GRU) networks with slight differences in structure or parameters, denoted as {GRU1, GRU2, …, GRU N}. The diversity among networks is guaranteed by using different random weight initialization or applying different dropout rates during training. At time t, the state vector X(t) is fed into all N GRU networks in parallel, and each network GRU i outputs an independent prediction sequence of the angle error in the future M time steps,

[0053]

[0054] Subsequently, an analysis module performs statistical processing on the N prediction sequences. For each future time point t + τ, the mean and variance of all prediction values are calculated. The prediction error mean μ e (t + τ) represents the consensus estimate of the most likely evolution trajectory of the future error, which will be used for prospective compensation control. The prediction error variance quantifies the degree of dispersion of all prediction trajectories, which serves as a key risk indicator and directly reflects the system's confidence in the current prediction. The larger the variance, the more serious the divergence of each network model on the future trend, and the more uncertain or dangerous the working conditions the system faces.

[0055] The third step is to generate a dual-channel adaptive risk control torque. Based on the MPE and VPE calculated by the analysis module, the main controller generates a main control torque. This torque is synthesized by two parallel channels.

[0056] The forward-looking compensation channel uses MPE to actively compensate for the error before it is fully formed. It calculates a feedforward compensation torque T ff (t). The magnitude of this moment is related to a key future time point t+τ c The MPE value is proportional to the time point τ c The selection is determined by the inherent response delay of the system.

[0057] The risk adaptive feedback channel is based on the current actual error e d (t) performs feedback control, but its behavior will be intelligently adjusted according to VPE. In this embodiment, a PID controller is used, and its output feedback torque is T fb (t). Its proportional gain K p , integral gain K i and differential gain K d are all functions of VPE. Specifically, when When it increases, it indicates that the system uncertainty increases. At this time, the differential gain K d will increase accordingly to enhance the damping and stability of the system, and the proportional gain K p and integral gain K i It will be appropriately reduced to avoid excessive control behavior under uncertain working conditions, thereby ensuring the driving safety of the vehicle.

[0058] The fourth step is active detection and online model correction based on uncertainty driving. This is the core link that distinguishes this invention from existing technologies. It gives the system the ability to self-learn and adapt when facing unknown environments.

[0059] A correction module continuously monitors the VPE value output by the analysis module. When VPE reaches a specific future time point t+τ r The value of exceeds a preset activation threshold Θ uc , indicating that the ETEP prediction model can no longer accurately describe the current vehicle dynamics, and the correction module is activated.

[0060] After activation, the correction module will generate a preset high-frequency, small-amplitude sinusoidal disturbance torque signal as the active detection signal T probe (t). The frequency and amplitude of this signal are carefully designed to effectively stimulate the system's dynamic response without being perceived by the driver or affecting normal driving safety. This detection signal is superimposed on the main control torque and injected into the steering actuation system.

[0061] At the same time, the correction module will accurately measure the actual response change of the vehicle caused by the probe signal, such as the slight fluctuation of yaw rate Δω z,act It will also use each GRU network in ETEP to predict the theoretical response Δω z,pred,i of the system that should be generated under the same probe signal. By comparing the actual response with the theoretical response of each model, the mismatch error ε

[0062] ε m,i of each model can be calculated. z,act z,pred,i

[0063] This mismatch error directly exposes the degree of mismatch of the current vehicle model. The correction module will immediately fine-tune the internal weights W i of all members in the ETEP integrated network in real time, small steps, according to this mismatch error. This process enables the internal models of ETEP to quickly approach the current real physical world, thereby rapidly reducing the divergence of ETEP internal prediction, so that the VPE value decreases. When the VPE falls below the threshold Θ uc , the active probe process stops, and the system returns to the regular high-confidence operation mode.

[0064] Fifth, the synthesis and execution of the total control torque. Finally, the total torque T out (t) output to the steering execution motor is synthesized by the torque generated by the main control module and the active probe signal. If the correction module is not activated, the total torque is the main control torque. This total torque signal acts on the steering execution motor through the driver to complete a closed-loop angle error correction. Subsequently, the system enters the next control period and repeats all the above steps.

[0065] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.​​

Claims

1. A wire-controlled steering angle error correction system based on neural network compensation, characterized in that: include: An acquisition module is used to obtain the real-time state vector of the vehicle; a prediction module, connected to the acquisition module, configured to input the real-time state vector into a timing error prediction integrated network, the timing error prediction integrated network being composed of a plurality of independent timing neural networks and outputting a plurality of independent predictions of future angular errors; an analysis module, connected to the prediction module, for calculating a prediction error mean and a prediction error variance of the future angular error based on the multiple independent predictions, wherein the prediction error variance is used to quantify the uncertainty of the prediction; a main control module, connected to the analysis module, and configured to generate a main control torque according to the prediction error mean and the prediction error variance; A correction module is connected to the analysis module and the prediction module. The correction module is used to determine whether the prediction error variance exceeds a preset activation threshold. If so, an active detection signal is generated and injected into the wire-controlled steering system, and the timing error prediction integrated network is corrected online based on the actual response of the wire-controlled steering system to the active detection signal.

2. The system according to claim 1, wherein: The step of the main control module generating the main control torque includes: calculating a forward-looking compensation torque based on the prediction error mean value, so as to compensate for the prediction error in advance.

3. The system according to claim 1, wherein: The step of the main control module generating the main control torque further includes: generating a risk-adaptive feedback torque through a feedback controller, wherein the core parameters of the feedback controller are dynamically adjusted in real time according to the prediction error variance.

4. The system according to claim 1, wherein: The analysis module calculates the prediction error variance The way is: Wherein, N is the number of the temporal neural networks, is the predicted value of the i-th time series neural network at the future time t+τ, μ e (t+τ) is the mean of the multiple independent predictions at time t+τ.

5. The system according to claim 1, wherein: The active detection signal is a preset high-frequency, small-amplitude disturbance signal, the frequency and amplitude of which are set to be imperceptible to the driver and not affect driving safety.

6. The system according to claim 1, wherein: The step of the correction module online correcting the timing error prediction integrated network includes: using each timing neural network in the timing error prediction integrated network to predict the theoretical response caused by the active detection signal, calculating the model mismatch error between the theoretical response and the actual response, and updating the weight of each timing neural network in the timing error prediction integrated network according to the model mismatch error.

7. The system according to claim 6, characterized in that The updating rule of the weight is: Among them, W i (t) is the weight of the i-th temporal neural network at time t, η is the learning rate, and L is the model mismatch error ε m,i The loss function of (t), is the gradient of the weight.

8. The system according to claim 1, wherein: The real-time state vector includes at least one of a command steering angle, an actual steering angle, a current angle error, a vehicle speed, and a yaw rate, or a combination of multiple items.

9. The system according to claim 1, wherein: The temporal neural network is a gated recurrent unit network or a long short-term memory network.

10. The system according to claim 1, wherein: The total torque ultimately output to the steering actuator is formed by the superposition of the main control torque and the active detection signal.

Images