An eps torque interface back correction method and system under intelligent driving mode

Abstract

The application discloses an EPS torque interface back-to-normal compensation method and system in an intelligent driving mode, and relates to the technical field of intelligent driving steering execution. The method comprises the following steps: acquiring a host computer torque request signal, a vehicle speed and a steering wheel rotation angle; performing filtering processing on the torque request signal, and performing direction judgment on the filtered torque request signal by using a lightweight time sequence multi-scale discrimination network; obtaining a basic back-to-normal torque in a manual driving condition by a calibration lookup table according to the vehicle speed and the steering wheel rotation angle; and performing dynamic correction on the basic back-to-normal torque according to a direction judgment result to obtain a torque compensation result. The application can accurately compensate the back-to-normal force of the EPS torque interface in the intelligent driving mode, and realizes stable control of the steering system.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving steering execution technology, and in particular to an EPS torque interface return compensation method and system in intelligent driving mode. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] When the vehicle turns, the driver releases the steering wheel, and the steering system's return-to-center module calculates and provides assistance based on parameters such as vehicle speed and steering wheel rotation speed. This assists the vehicle's suspension in generating its own return-to-center force, allowing the steering wheel to return precisely to the center position. When the vehicle's intelligent driving function is activated, the steering wheel is taken over by the intelligent driving function. In the torque interface solution, because the existing return-to-center module of the steering system cannot recognize the direction of the torque requested by the host computer, the applied return-to-center torque compensation may be contrary to the host computer's execution intention, thus interfering with the normal execution of the intelligent driving function and failing to achieve the return-to-center calibration of the intelligent driving function. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide an EPS torque interface return-to-center compensation method and system in intelligent driving mode, which accurately compensates for the return-to-center force of the EPS torque interface in intelligent driving mode, thereby achieving stable control of the steering system.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a method for EPS torque interface return compensation in intelligent driving mode, comprising the following steps: Acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations; The torque request signal is filtered, and a lightweight temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by calibrating and looking up a table. The basic return torque is dynamically corrected based on the direction determination result to obtain the torque compensation result.

[0006] Further preprocessing operations include outlier removal from the torque signal and linear interpolation to complete missing values ​​of the host computer torque request signal, vehicle speed, and steering wheel angle.

[0007] Furthermore, an adaptive first-order low-pass filter is used to filter the torque request signal, and the filter coefficients are dynamically adjusted according to the current torque gradient and vehicle speed.

[0008] Furthermore, the specific steps for determining the direction of the filtered torque request signal using a lightweight temporal multi-scale discriminant network are as follows: Design a lightweight temporal multi-scale discriminant network; Construct a sample set using known data and perform automatic annotation and data augmentation on the samples; A lightweight temporal multi-scale discriminant network is trained using a sample set; The trained temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal.

[0009] Furthermore, a lightweight temporal multi-scale discriminant network is trained by introducing a modulation factor using a focus loss function.

[0010] Furthermore, the specific steps for dynamically correcting the base return torque based on the direction determination result are as follows: When at the rising edge, the return-to-center compensation intensity is reduced to avoid counteracting the driver's intention to actively steer; when at the falling edge, the return-to-center compensation intensity is increased to assist the steering wheel in returning to center; when in a stable holding state, the basic return-to-center torque remains unchanged.

[0011] A second aspect of the present invention provides an EPS torque interface return-to-center compensation system in intelligent driving mode, comprising: The data acquisition module is configured to acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations. The direction determination module is configured to filter the torque request signal and use a lightweight temporal multi-scale discriminant network to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. The torque preliminary judgment module is configured to obtain the basic return torque under manual driving conditions by looking up a table based on the vehicle speed and steering wheel angle. The dynamic correction module is configured to dynamically correct the basic return torque based on the direction judgment result, and obtain the torque compensation result.

[0012] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute steps in the EPS torque interface return compensation method in intelligent driving mode as described in the first aspect of the present invention.

[0013] A fourth aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the EPS torque interface return compensation method in intelligent driving mode as described in the first aspect of the present invention.

[0014] A fifth aspect of the present invention provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the EPS torque interface return-to-center compensation method in intelligent driving mode as described in the first aspect of the present invention.

[0015] The above one or more technical solutions have the following beneficial effects: This invention discloses an EPS torque interface return-to-center compensation method and system in intelligent driving mode. Through a return-to-center scheme based on calibrable parameters corrected by deep learning, when the host computer requests steering wheel angle correction, the steering system applies a correct return-to-center torque, ensuring the steering wheel angle accurately returns to the center position and the vehicle trajectory matches the calibration expectation. This results in better performance in lane keeping and adaptive cruise control functions, reducing problems such as lane departures and lane changes caused by slow steering response.

[0016] This invention improves the smoothness of the torque curve through adaptive filtering, uses a deep learning model to replace the instantaneous gradient, reduces the direction misjudgment rate in sensor noise environments, enhances the recognition accuracy of boundary ambiguity areas (slow torque changes or small fluctuations), and makes direction judgment more timely and accurate.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 Schematic diagram of the return-to-center compensation principle for the rectangular interface; Figure 2 This is a schematic diagram illustrating the changes in torque and gradient over time. Figure 3 This is a flowchart of the EPS torque interface return compensation method in the intelligent driving mode of Embodiment 1 of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] In Advanced Driver Assistance Systems (ADAS) driving functions, the control interface for the Electric Power Steering (EPS) system typically takes two forms: torque interface and angle interface. In the torque interface form, the EPS acts as an actuator, converting the torque request signal input from the host computer into a corresponding motor current output torque. The principle is as follows... Figure 1 As shown, ADAS sends a torque request signal to the torque controller to generate a filtered torque Tb. The filtered torque is then compensated by the homing module to obtain the homing compensation torque Td. Td and Tb are superimposed to form the output motor torque.

[0023] like Figure 2As shown, the slope of the requested torque curve at various points indicates, to some extent, the speed at which the host computer wants to adjust the vehicle's attitude, and can be used as a parameter for calibrating the magnitude of the return torque. Steering wheel torque is typically defined as positive on the left and negative on the right, with one cycle consisting of turning the wheel fully to the left or right and returning it to the center position. A complete cycle is divided into four intervals. When the product of torque and gradient is positive, the requested torque is on the rising edge, and the return torque compensation is 0. When the product of torque and gradient is negative, the requested torque is on the falling edge, and the return torque needs to be applied.

[0024] The existing EPS return-center module cannot recognize the direction of the torque requested by the host computer. The applied return-center torque compensation may be contrary to the host computer's execution intention, thus interfering with the normal execution of ADAS functions. Therefore, when ADAS driving functions are enabled, the EPS return-center module cannot be directly invoked and is usually in a suppressed state.

[0025] When the host computer requests the steering wheel angle to be adjusted back, but the requested torque is in the pullback phase and has not yet reversed direction, the direction of the EPS motor output torque does not change. At this time, the vehicle relies on the self-centering force generated by the chassis's own mechanical structure to overcome various resistances (including road friction, internal friction of the steering system, etc.). If the vehicle system friction or other road factors cause the self-centering resistance to be inconsistent with the expected working conditions during calibration, the host computer's torque pullback gradient will be relatively small and will not be able to meet the host computer's need to adjust the vehicle's attitude in time. This will easily cause problems such as crossing the line or making a "dragon's tail" in cornering and lane changing scenarios.

[0026] To address the aforementioned problems, this invention discloses a method and system for EPS torque interface return compensation in intelligent driving mode, the details of which are described in the following embodiments.

[0027] Example 1: Embodiment 1 of the present invention provides a method for EPS torque interface return compensation in intelligent driving mode, such as... Figure 3 As shown, it includes the following steps: S1: Obtain the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations.

[0028] In one specific implementation, the torque request signal uses a 1000 Hz sampling frequency to ensure detail capture, and the vehicle speed and steering angle signals are collected at their respective original frequencies and then synchronized to the same time base through interpolation.

[0029] Preprocessing includes outlier removal from the torque signal and linear interpolation to complete missing values ​​in the host computer's torque request signal, vehicle speed, and steering wheel angle. Specifically, a sliding window is used to determine if the torque value exceeds the reasonable range set by the engineering; if it is abnormal, the previous valid value is used as the replacement. Simultaneously, linear interpolation is performed to complete missing values ​​to ensure signal continuity.

[0030] S2: Filter the torque request signal and use a lightweight time-series multi-scale discriminant network to determine the direction of the filtered torque request signal.

[0031] S2.1: Filter the torque request signal.

[0032] In one specific implementation, to eliminate high-frequency sensor noise, the torque request signal from the host computer first needs to be processed to eliminate high-frequency oscillations and make the curve smooth and flat. This facilitates subsequent calculations and prevents the alignment module from frequently opening and closing, thus losing its effectiveness. This embodiment uses an adaptive first-order low-pass filter to filter the torque request signal. The filter coefficient is dynamically adjusted according to the current torque gradient and vehicle speed, aiming to eliminate high-frequency oscillations and make the curve smooth and flat, while avoiding excessive delay in response. The baseline value of the filter coefficient is determined by the sampling frequency and the target cutoff frequency, and its value ranges from zero to one. When the torque gradient is large and the vehicle speed is low, the filter coefficient is reduced to enhance the smoothing effect and prevent the subsequent alignment module from frequently opening and closing due to noise; when the gradient is small or the vehicle speed is high, the filter coefficient is increased to ensure response speed. Finally, the filtered torque and its historical buffer value are output.

[0033] Specifically, in this embodiment, a first-order low-pass hardware or software filter is used, with the cutoff frequency set to 100 Hz. All signals are timestamped and then downsampled to 100 Hz to balance real-time performance and computational load. The formula is as follows: .

[0034] in, This is the filtered torque request signal. The current torque value requested by the host computer. This is the result of the previous filtering (the initial filtering settings are based on experience). These are the filter coefficients, with values ​​between 0 and 1.

[0035] S2.2: Use a lightweight temporal multi-scale discriminant network to determine the direction of the filtered torque request signal.

[0036] S2.2.1: Design a lightweight temporal multi-scale discriminant network.

[0037] In one specific implementation, the temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from the filtered torque request signal from different time scales. The second layer is used to evaluate the importance of the scales based on the extracted features and reallocate the weights. The third layer is used to determine the key change moments in the torque curve.

[0038] Specifically, the first layer is a parallel multi-scale convolution module. This module breaks the limitation of the traditional single convolution kernel size and uses three different length convolution kernels to process the input torque sequence in parallel. The short convolution kernel has a receptive field of 3 sampling points, which is good at capturing transient changes in torque; the medium convolution kernel has a receptive field of 7 sampling points, which is suitable for recognizing slow trends lasting for tens of milliseconds; and the long convolution kernel has a receptive field of 15 sampling points, which can perceive the overall trend of the torque curve.

[0039] These three parallel convolutional layers compute independently, generating three sets of feature maps, each containing eight feature channels. The three sets of features are then concatenated along the channel dimension to form a comprehensive feature representation with 24 channels. The advantage of this design is that the network can extract appropriate features regardless of whether the torque change is rapid or gradual.

[0040] The second layer is the channel adaptive recalibration module. This is an improved compressed excitation module for time-series signals. Traditional compressed excitation modules are used for image recognition; this embodiment adapts it to one-dimensional time-series data. This module first performs a global compression operation on each feature channel, calculating the average value of that channel over 64 time steps, compressing the 24 channels into a 24-dimensional vector, where each dimension represents the overall activation intensity of the corresponding channel. Then, it is excited through two fully connected network layers: the first layer compresses the 24 dimensions to 6 dimensions by introducing a non-linear transformation; the second layer restores the 6 dimensions to 24 dimensions and uses the sigmoid function to restrict the output to between 0 and 1, obtaining the weight coefficients for each channel. Finally, the original 24-channel feature map is multiplied channel by channel with these weights to achieve feature recalibration.

[0041] The core value of this mechanism lies in the fact that when the torque is at the rising edge, the network automatically assigns higher weights to short convolutional kernel features; when the torque changes gradually, long convolutional kernel features receive greater weights; thus achieving dynamic adaptation of scale selection.

[0042] The third layer is the temporal attention focusing module. Its purpose is to replace simple global averaging or last-moment extraction. This module learns an attention weight distribution related to time steps and automatically determines which moments in the 64 time steps are most important for determining the direction.

[0043] The specific implementation involves using a one-dimensional convolutional layer with a kernel size of 1 to map the 24-channel features into single-channel attention scores, with one score corresponding to each time step. Then, a Softmax function is applied to these scores to normalize them into a probability distribution, with a sum of 100. Finally, the original 24-channel features are weighted and summed with these attention weights to obtain the final 24-dimensional feature vector.

[0044] For example, near the inflection point where torque changes from increasing to decreasing, the network automatically assigns higher weights to the time steps before and after the inflection point because the slope change information is richest at these moments.

[0045] Finally, the weighted summation of the 24-dimensional feature vector is fed into the classification head. The classification head contains two fully connected layers. The first layer maps the 24 dimensions to 12 dimensions and introduces random deactivation to prevent overfitting. The second layer compresses the 12 dimensions to 3 dimensions, corresponding to the rising edge, falling edge, and stationary states. Finally, the probability distribution of the three classes is output through the Softmax function.

[0046] S2.2.2: Construct a sample set using known data and perform automatic labeling and data augmentation on the samples.

[0047] In one specific implementation, a known filtered torque request sequence is obtained as a sample set from real vehicle data or a high-precision simulation platform, with a sampling frequency of 1000 Hz. Vehicle speed and steering wheel angle are simultaneously collected as auxiliary references. Data collection covers various driving scenarios: frequent low-speed steering in urban areas, minor corrections at high speeds, sharp steering for emergency avoidance, and noise interference on bumpy roads, etc.

[0048] The sample set includes: Rising edge sample: refers to the time period during which the torque request shows a continuous increasing trend. During this time, the driver is actively increasing the steering torque, such as initiating steering or further increasing the steering angle. Its characteristic is that the filtered torque increases over time with a positive gradient, and the duration typically exceeds 50 milliseconds.

[0049] Falling edge sample: refers to the period during which the torque request is continuously decreasing. At this time, the torque is decreasing, possibly due to the driver actively returning the steering wheel to center or the steering wheel naturally returning to center. Its characteristic is that the torque decreases over time, the gradient is negative, and the duration exceeds 50 milliseconds.

[0050] Stable holding period: refers to the time period during which torque request does not show a significant upward or downward trend. During this period, the torque is basically stable and may be in a state of standstill, uniform steering, or slight fluctuation. It is characterized by a small absolute value of the gradient or a change duration of less than 50 milliseconds.

[0051] To reduce computational burden, the signal is downsampled to 100 Hz after acquisition, using an average of 10 consecutive sampling points to reduce frequency while preserving signal characteristics. After downsampling, a sample window is defined as 64 time steps, with adjacent windows overlapping by 50 time steps to ensure temporal continuity.

[0052] Due to the extremely high cost of manual annotation, an automatic annotation method based on rules and validation is designed. Initial labels are generated using a combination of rules based on the original torque value and gradient, followed by post-processing optimization. Specifically, for each sampling point, the average slope of the sliding window of the filtered torque is calculated, with a window length of 5 sampling points, or 50 milliseconds. When the average slope is greater than 0.5 Nm / s and its duration exceeds 50 milliseconds, the central interval of that time period is marked as a rising edge; when the average slope is less than -0.5 Nm / s and its duration exceeds 50 milliseconds, it is marked as a falling edge; all other cases are marked as stable.

[0053] To avoid frequent state transitions, a protective interval mechanism is added. A 20-millisecond transition zone is set before and after each state boundary; this region is not included in training because the true state at the boundary is inherently ambiguous. Finally, 1% of the samples are manually sampled for correction to ensure labeling quality.

[0054] To address data imbalance and improve generalization ability, online data augmentation is performed on the samples before training. This embodiment designs three augmentation methods, and one or a combination of them is randomly selected for each training session.

[0055] The first method is temporal stretching and compression, which randomly stretches or compresses the time axis of the entire 64-time-step sequence by no more than 10% to simulate the differences in operating speed under different driving styles. Linear interpolation is used during stretching, and uniform sampling is used during compression.

[0056] The second method involves adding Gaussian white noise, with the signal-to-noise ratio controlled between 30 and 40 dB, to simulate the electronic noise of real sensors and data acquisition links. The noise amplitude is consistent with the actual system's measured noise level.

[0057] The third method is amplitude jitter, which involves multiplying the entire torque sequence by a random coefficient that is uniformly distributed between 0.95 and 1.05 to simulate the gain deviation of torque sensors between different vehicles.

[0058] S2.2.3: Train a lightweight temporal multi-scale discriminant network using the sample set.

[0059] In one specific implementation, this embodiment uses a focal loss function instead of standard cross-entropy to address the class imbalance problem where there are far more stationary state samples than rising and falling edge samples. Specifically, the focal loss function is used to train a lightweight temporal multi-scale discriminant network by introducing a modulation factor, allowing samples that are more difficult for the model to train to receive higher weights, namely those rising and falling edge boundary samples that are easily misclassified.

[0060] In torque direction determination tasks, the number of samples for rising and falling edges is far less than that for stationary states, resulting in a severe class imbalance problem. Standard cross-entropy loss tends to cause the model to predict stationary states with a large number of samples, ignoring the rare edge state samples. The focus loss function addresses this problem by introducing a modulation factor. For each sample, the difference between the predicted probability and the true label is calculated. When the sample is easily classified correctly, the modulation factor approaches zero, significantly reducing the loss; when the sample is difficult to classify correctly, the modulation factor is larger, preserving the loss. Simultaneously, a class weight coefficient is introduced to assign higher weights to rising and falling edges.

[0061] This mechanism forces the model to shift its training focus from a large number of simple, stable samples to a small number of difficult, along-the-line samples, making the network pay more attention to the rising and falling transition regions of the torque curve, thereby significantly improving the accuracy of identifying rare work states.

[0062] An early stopping strategy is employed during training: training is halted when the validation set loss does not decrease for 20 consecutive epochs to prevent overfitting. Simultaneously, precision and recall for each class are monitored to ensure balanced performance across the three classes without sacrificing the minority class.

[0063] S2.2.4: Use the trained temporal multi-scale discriminant network to determine the direction of the filtered torque request signal.

[0064] In one specific implementation, after the model is deployed to the vehicle controller, a torque history buffer of fixed length 64 is maintained and updated every 10 milliseconds according to the sampling frequency. Each time a new torque value arrives, the entire buffer is fed into the model for one inference.

[0065] To eliminate jitter in a single inference, probability smoothing is performed on the results of multiple consecutive inferences. Specifically, the three probabilities output by the current model are weighted and averaged with the previously smoothed probabilities in a 7:3 ratio. The state is switched only when the maximum smoothed probability exceeds a threshold of 0.7; otherwise, the previous stable state is maintained. This mechanism avoids frequent jumps in the output state.

[0066] After the model outputs three probability values, they need to be mapped to the gradient sign variable in the original algorithm to ensure compatibility with the downstream positive feedback module. When the rising edge probability exceeds 0.7, a positive one is output to indicate that the current trend is upward; when the falling edge probability exceeds 0.7, a negative one is output to indicate that the trend is downward; otherwise, a zero is output to indicate that the torque is stable.

[0067] At the same time, the model also outputs three original probability values ​​for diagnostic records, which facilitates subsequent analysis of the rationality of the model's behavior.

[0068] After quantization and compression, the model requires less than 20KB of storage. On a typical automotive microcontroller, a single inference operation takes approximately 0.8 to 1.2 milliseconds, far less than the 10-millisecond sampling period, meeting real-time requirements. Inference computation is primarily consumed by parallel convolution operations, but due to the small size of the convolution kernels and the limited number of channels, the computational load remains within acceptable limits.

[0069] In this embodiment, during the direction determination of the filtered torque request signal using a trained temporal multi-scale discriminant network, a 64-bit buffer of filtered torque history is maintained, and inference is performed every 10 milliseconds. Internally, the network first extracts features from three time scales simultaneously through a parallel multi-scale convolutional module: a short convolutional kernel with a receptive field of 3 sampling points captures transient torque jumps; a medium convolutional kernel with a receptive field of 7 sampling points identifies continuous, gradual trends; and a long convolutional kernel with a receptive field of 15 sampling points perceives the overall direction. These three sets of features are concatenated to obtain a 24-channel comprehensive feature representation.

[0070] Subsequently, the importance weight of each feature channel is calculated through a channel adaptive recalibration module. This module calculates the global average value on the time axis for each of the 24 channels, and outputs a weight coefficient between 0 and 1 after compression excitation, which is then used to weight the original features. This allows the network to automatically adjust its scale preference according to the current torque change rate, emphasizing short-term features during rising edges and long-term features during gradual changes.

[0071] Next, the temporal attention focusing module is used to learn the attention distribution over 64 time steps, automatically determining which moments are most critical for direction judgment. After weighted summation of attention weights, a 24-dimensional feature vector is obtained, which is fed into the classification head to output the probabilities of three states: rising edge, falling edge, and stable hold.

[0072] To avoid jitter in a single inference iteration, continuous outputs are probabilistically smoothed by averaging the current probability with the historical smoothed probability in a 7:3 ratio. A new state is output only when the smoothed maximum probability exceeds 0.7; otherwise, the previous state is retained. Finally, the output is mapped back to the gradient sign variable of the traditional algorithm, with rising edges corresponding to positive gradient values ​​and falling edges to negative gradient values, smoothly maintaining the corresponding zero value.

[0073] It should be noted that, in addition to step S2, this embodiment can also use the instantaneous gradient calculation method for direction determination. The specific method is as follows: Calculate the torque gradient between two adjacent frames of the filtered signal: .

[0074] in, This indicates the current filtering result. For time intervals.

[0075] Direction of torque acquisition :when hour, ;when hour, ; Torque gradient relative to torque direction Multiply to confirm whether the current torque request is at the rising or falling edge.

[0076] .

[0077] When torque request value If negative, the host computer requests torque at the falling edge; when If the value is positive, the host computer's torque request is on the rising edge.

[0078] S3: Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by looking up the table through calibration.

[0079] In one specific implementation, the return-to-center parameters are calibrated and determined based on the vehicle speed and steering wheel angle, as shown in Table 1, and the basic return-to-center torque under manual driving conditions is calculated. This step can directly call the result calculated by the return-to-center module under normal operating conditions.

[0080] Table 1. Calibration and determination of return-to-center parameters

[0081] S4: Dynamically correct the basic return torque based on the direction judgment result to obtain the torque compensation result.

[0082] In one specific implementation, this embodiment adjusts the gain of the base return torque based on the direction judgment result output by the deep learning model. The direction judgment result includes probability values ​​in three categories: rising edge probability, falling edge probability, and stable holding probability. When it is at the rising edge, the return torque compensation intensity is reduced to avoid counteracting the driver's active steering intention; when it is at the falling edge, the return torque compensation intensity is increased to assist the steering wheel in returning to center; when it is in a stable holding state, the base return torque remains unchanged. Simultaneously, vehicle speed and steering angle confidence are weighted, and when the vehicle speed is extremely low or the steering angle is very small, the return torque is linearly reduced to zero to prevent accidental triggering while stationary.

[0083] The system uses a combination of factors, including the state output by the deep learning model, vehicle speed, steering angle, and absolute torque value, to determine whether return-to-center intervention is allowed. Allowed return-to-center intervention is permitted when the vehicle speed is above the minimum return-to-center speed threshold, the steering wheel angle is greater than the minimum return-to-center angle, the vehicle is currently in a falling edge or stable holding state, and there are no sensor malfunctions. Prohibited return-to-center intervention is prohibited when the vehicle is in a rising edge, the torque request changes too drastically, or the absolute torque value is too low.

[0084] The filtered torque request is added to the gain-corrected return torque to obtain the final output torque to the actuator. The return engagement coefficient can be gradually increased according to the duration of the falling edge to achieve smooth engagement rather than abrupt change. Before output, it undergoes dual amplitude limiting protection of absolute value and gradient to prevent sudden torque changes from causing steering wheel vibration or impact.

[0085] The final torque command is sent to the motor drive unit at 5-millisecond intervals. Simultaneously, key diagnostic signals are recorded for online monitoring and offline analysis, including the filtered torque value, the three probabilities output by the deep learning model, the current state determination result, and the actual contribution value of the homing compensation. This data is output through the onboard diagnostic interface for subsequent calibration optimization and fault analysis.

[0086] It should be noted that, in addition to step S4, the torque compensation result can also be determined in the following manner in this embodiment: Calculate the angle θ of the tangent line to the filtered torque curve: .

[0087] The value in arrive Between, through actual vehicle calibration With compensation coefficient The correspondence is calculated using interpolation. When hour, All The parameter table is shown in Table 2.

[0088] Table 2. Vehicle Calibration With compensation coefficient Correspondence parameters

[0089] Final output return torque : .

[0090] when When the value is always 1, the return torque applied by the ADAS module is the same as the return torque under normal operating conditions. At this time, the torque request processing of the host computer is closest to the hand torque of the driver in actual vehicle operation.

[0091] Example 2: Embodiment 2 of the present invention provides an EPS torque interface return-to-center compensation system in intelligent driving mode, comprising: The data acquisition module is configured to acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations. The direction determination module is configured to filter the torque request signal and use a lightweight temporal multi-scale discriminant network to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. The torque preliminary judgment module is configured to obtain the basic return torque under manual driving conditions by looking up a table based on the vehicle speed and steering wheel angle. The dynamic correction module is configured to dynamically correct the basic return torque based on the direction judgment result, and obtain the torque compensation result.

[0092] Example 3: Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted for loading and execution by a processor of the steps in an EPS torque interface return-to-center compensation method in intelligent driving mode as described in Embodiment 1 of the present invention, wherein the steps are: Acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations; The torque request signal is filtered, and a lightweight temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by calibrating and looking up a table. The basic return torque is dynamically corrected based on the direction determination result to obtain the torque compensation result.

[0093] The detailed steps are the same as those provided in Embodiment 1 for EPS torque interface return compensation in intelligent driving mode, and will not be repeated here.

[0094] Example 4: Embodiment 4 of the present invention provides a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps of an EPS torque interface return-to-center compensation method in intelligent driving mode as described in Embodiment 1 of the present invention, wherein the steps are: Acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations; The torque request signal is filtered, and a lightweight temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by calibrating and looking up a table. The basic return torque is dynamically corrected based on the direction determination result to obtain the torque compensation result.

[0095] The detailed steps are the same as those provided in Embodiment 1 for EPS torque interface return compensation in intelligent driving mode, and will not be repeated here.

[0096] Example 5: Embodiment 5 of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the EPS torque interface return-to-center compensation method in intelligent driving mode as described in Embodiment 1 of the present invention. The steps are as follows: Acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations; The torque request signal is filtered, and a lightweight temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by calibrating and looking up a table. The basic return torque is dynamically corrected based on the direction determination result to obtain the torque compensation result.

[0097] The detailed steps are the same as those provided in Embodiment 1 for EPS torque interface return compensation in intelligent driving mode, and will not be repeated here.

[0098] The steps and methods involved in Examples 2, 3, 4 and 5 above correspond to those in Example 1. For specific implementation methods, please refer to the relevant description section of Example 1.

[0099] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0100] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for EPS torque interface return compensation in intelligent driving mode, characterized in that, Includes the following steps: Acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations; The torque request signal is filtered, and a lightweight temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. Based on the vehicle speed and steering wheel angle, the basic return torque under manual driving conditions is obtained by calibrating and looking up a table. The basic return torque is dynamically corrected based on the direction determination result to obtain the torque compensation result.

2. The EPS torque interface return-to-center compensation method in intelligent driving mode as described in claim 1, characterized in that, The preprocessing operations include outlier removal from the torque signal and linear interpolation to complete missing values ​​of the host computer torque request signal, vehicle speed, and steering wheel angle.

3. The EPS torque interface return-to-center compensation method in intelligent driving mode as described in claim 1, characterized in that, An adaptive first-order low-pass filter is used to filter the torque request signal, and the filter coefficients are dynamically adjusted according to the current torque gradient and vehicle speed.

4. The EPS torque interface return-to-center compensation method in intelligent driving mode as described in claim 1, characterized in that, The specific steps for determining the direction of the filtered torque request signal using a lightweight temporal multi-scale discriminant network are as follows: Design a lightweight temporal multi-scale discriminant network; Construct a sample set using known data and perform automatic annotation and data augmentation on the samples; A lightweight temporal multi-scale discriminant network is trained using a sample set; The trained temporal multi-scale discriminant network is used to determine the direction of the filtered torque request signal.

5. The EPS torque interface return-to-center compensation method in intelligent driving mode as described in claim 4, characterized in that, A lightweight temporal multi-scale discriminant network is trained by using a focus loss function and introducing a modulation factor.

6. The EPS torque interface return-to-center compensation method in intelligent driving mode as described in claim 1, characterized in that, The specific steps for dynamically correcting the base return torque based on the direction determination result are as follows: When at the rising edge, the return-to-center compensation intensity is reduced to avoid counteracting the driver's intention to actively steer; when at the falling edge, the return-to-center compensation intensity is increased to assist the steering wheel in returning to center; when in a stable holding state, the basic return-to-center torque remains unchanged.

7. An EPS torque interface return-to-center compensation system in intelligent driving mode, characterized in that, include: The data acquisition module is configured to acquire the torque request signal from the host computer, the vehicle speed, and the steering wheel angle, and perform preprocessing operations. The direction determination module is configured to filter the torque request signal and use a lightweight temporal multi-scale discriminant network to determine the direction of the filtered torque request signal. The temporal multi-scale discriminant network adopts a three-layer cascaded processing architecture. The first layer is used to extract features from different time scales of the filtered torque request signal. The second layer is used to evaluate the importance of the scales and reallocate weights based on the extracted features. The third layer is used to determine the key change moments in the torque curve. The torque preliminary judgment module is configured to obtain the basic return torque under manual driving conditions by looking up a table based on the vehicle speed and steering wheel angle. The dynamic correction module is configured to dynamically correct the basic return torque based on the direction judgment result, and obtain the torque compensation result.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the EPS torque interface return compensation method in intelligent driving mode as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-6, for the EPS torque interface return compensation method in intelligent driving mode.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the EPS torque interface return compensation method in intelligent driving mode as described in any one of claims 1-6.

Images