Rolling force prediction method and related methods, apparatuses, devices, and media

Abstract

This application discloses a rolling force prediction method and related methods, apparatus, devices, and media. The rolling force prediction method includes: extracting features based on sensor data at a first time moment to obtain a feature manifold at the first time moment; analyzing the feature manifold at the first time moment and the feature manifold at a second time moment prior to the first time moment to obtain the sliding variance of the feature manifold within a target sliding window; determining a first value of a first weight at the first time moment based on the sliding variance, and determining a second value of a second weight at the first time moment based on the first value and a negative correlation; wherein the sliding variance is negatively correlated with the first weight; and predicting the feature manifold at the first time moment based on a rolling force prediction model to obtain the predicted rolling force at the first time moment. This scheme can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a rolling force prediction method and related methods, apparatus, devices and media. Background Technology

[0002] In autonomous driving systems, rolling force is a core physical quantity that affects energy management and trajectory tracking accuracy. Therefore, it is necessary to predict rolling force as accurately as possible in order to better achieve autonomous driving.

[0003] Current technologies rely on fixed empirical formulas to model rolling forces. However, due to the high uncertainty of road conditions, tire pressure, and load distribution, these technologies often suffer from model mismatch or response lag when facing complex operating conditions. While machine learning methods such as support vector machines and artificial neural networks can handle nonlinear characteristics, their computational complexity is excessively high. Therefore, reducing the computational complexity and improving the accuracy of rolling force prediction has become an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a rolling force prediction method, as well as related methods, apparatus, equipment, and media, which can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction.

[0005] To address the aforementioned technical problems, the first aspect of this application provides a rolling force prediction method, comprising: extracting features based on sensor data at a first moment to obtain a feature manifold at the first moment; analyzing the feature manifold at the first moment and the feature manifold at a second moment prior to the first moment to obtain the sliding variance of the feature manifold within a target sliding window; wherein the target sliding window covers the first moment and the second moment prior to the first moment, and the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively, wherein the polynomial kernel is responsible for the long-term trend of the rolling force, the Gaussian kernel is responsible for capturing local instantaneous jumps, and the first weight and the second weight are negatively correlated; determining a first value of the first weight at the first moment based on the sliding variance, and determining a second value of the second weight at the first moment based on the first value and the negative correlation; wherein the sliding variance is negatively correlated with the first weight; and predicting the feature manifold at the first moment based on the rolling force prediction model to obtain the predicted rolling force at the first moment.

[0006] To address the aforementioned technical problems, a second aspect of this application provides a driving control method, comprising: predicting a predicted rolling force based on sensor data at the current moment; wherein the predicted rolling force at the current moment is obtained based on the rolling force prediction method of the first aspect described above; and performing torque control on the vehicle motor based on the predicted rolling force at the current moment.

[0007] To address the aforementioned technical problems, a third aspect of this application provides a rolling force prediction device, comprising: a feature extraction module, a manifold analysis module, a weight update module, and a rolling prediction module. The feature extraction module is used to extract features based on sensor data at a first time moment to obtain the feature manifold at that first time moment. The manifold analysis module is used to analyze the feature manifold at the first time moment and the feature manifold at a second time moment prior to the first time moment to obtain the sliding variance of the feature manifold within a target sliding window. The target sliding window covers the first time moment and the second time moment prior to the first time moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps. The first weight and the second weight are negatively correlated. The weight update module is used to determine a first value of the first weight at the first time moment based on the sliding variance, and to determine a second value of the second weight at the first time moment based on the first value and the negative correlation. The sliding variance is negatively correlated with the first weight. The rolling prediction module is used to predict the feature manifold at the first time moment based on the rolling force prediction model to obtain the predicted rolling force at the first time moment.

[0008] To address the aforementioned technical problems, a fourth aspect of this application provides a driving control device, comprising: a current prediction module and a torque control module. The current prediction module is used to predict the rolling force at the current moment based on sensor data at the current moment; wherein the predicted rolling force at the current moment is obtained based on the rolling force prediction device of the second aspect described above. The torque control module is used to perform torque control on the vehicle motor based on the predicted rolling force at the current moment.

[0009] To address the aforementioned technical problems, the fifth aspect of this application provides an electronic device comprising at least a memory and a processor coupled to each other. The memory stores at least program instructions, and the processor executes the program instructions to implement the rolling force prediction method of the first aspect or the driving control method of the second aspect.

[0010] To address the aforementioned technical problems, a sixth aspect of this application provides a computer-readable storage medium storing program instructions executable by a processor. These program instructions are used to implement the rolling force prediction method of the first aspect or the driving control method of the second aspect.

[0011] The above scheme extracts features from sensor data at the first moment to obtain the feature manifold at that moment. Then, it analyzes the feature manifold at the first moment and the feature manifold at the second moment prior to the first moment to obtain the sliding variance of the feature manifold within the target sliding window. The target sliding window covers the first moment and the second moment prior to the first moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of rolling force, while the Gaussian kernel is responsible for capturing local instantaneous jumps. The first and second weights are negatively correlated. Therefore, based on the sliding variance, the first value of the first weight at the first moment is determined, and based on the first value and the negative correlation, the second weight at the first moment is determined. The second value at time step 1 is used, and the sliding variance is negatively correlated with the first weight. Then, based on the rolling force prediction model, the characteristic manifold at time step 1 is predicted to obtain the predicted rolling force at time step 1. On the one hand, the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with the first and second weights respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps, which helps to maximize adaptability to complex working conditions. Furthermore, since only the characteristic manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, it also helps to reduce the computational complexity of rolling force prediction. On the other hand, the characteristic manifold extracted from sensor data is adaptively... The first value of the first weight and the second value of the second weight should be determined accordingly. Since the sliding variance of the characteristic manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can be determined which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the data feedback from the sensors fluctuates less, and the local variability of the characteristic manifold is low, resulting in a smaller sliding variance. Therefore, the first value of the first weight can be adaptively increased while the second value of the second weight can be decreased, allowing the rolling force prediction model to focus more on the polynomial kernel to accurately control the long-term trend. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensors fluctuates at high frequencies, and the special... As the eigenvariance increases rapidly, the adaptive weights react immediately. The first value of the first weight decreases adaptively, and the second value of the second weight increases adaptively. This allows the rolling force prediction model to switch to a model that focuses more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted to follow the actual working conditions, which helps improve the accuracy of rolling force prediction. On the other hand, when calculating the eigenvariance, the sliding variance is calculated within the target sliding window by combining the eigenmanifold at the first moment and the eigenmanifold at the second moment before the first moment. This can filter out instantaneous sampling noise and help ensure that the evolution of the weight factors is as smooth and continuous as possible between different working conditions, which also helps improve the accuracy of rolling force prediction.Therefore, it can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating an embodiment of the rolling force prediction method of this application; Figure 2 This is a flowchart illustrating an embodiment of the driving control method of this application; Figure 3 This is a schematic diagram of the framework of an embodiment of the training device for the rolling force prediction model of this application; Figure 4 This is a schematic diagram of the framework of an embodiment of the driving control device of this application; Figure 5 This is a schematic diagram of the framework of an embodiment of the electronic device of this application; Figure 6 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0013] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0014] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0015] In this paper, the terms "system" and "network" are often used interchangeably. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the slash " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this paper indicates two or more objects.

[0016] Please refer to the following: Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the rolling force prediction method of this application. It should be noted that the process operations in this embodiment can be executed by an electronic device with computing capabilities or related equipment containing an electronic device. The specific structure and type of the electronic device and related equipment containing an electronic device are not limited herein. Specifically, this embodiment may include the following steps: Step S11: Perform feature extraction based on the sensor data at the first moment to obtain the feature manifold at the first moment.

[0017] In one implementation scenario, the first moment can be any moment when the sensor acquires data. For example, the sensor can acquire data at time t, t+1, t+2, ..., t+N, so the first moment can be any of these times. In practical applications, each moment when the sensor acquires data can be selected sequentially as the first moment to obtain the predicted rolling force at each moment, facilitating subsequent driving control.

[0018] In one implementation scenario, sensor data may include sensing signals from several sensors, including but not limited to: accelerometer sensing signals (e.g., containing attitude information), gyroscope sensing signals (e.g., containing attitude information), wheel speed sensor sensing signals (e.g., containing real-time recorded rotation frequency of each wheel), camera sensing signals (e.g., containing images of the road surface and surrounding environment), LiDAR sensing signals (e.g., containing point cloud data of the road surface and surrounding environment), etc., without limitation. Of course, the above examples are merely one possible instance of sensor data in practical applications, and other possible scenarios are not limited here, nor will they be listed one by one. It should be noted that in this embodiment, sensor data can be heterogeneous data (e.g., containing multiple sensing signals that differ in physical meaning, sampling frequency, etc.). Of course, the possibility of homogeneous sensor data is not excluded, and the specific type of sensor data is not limited here.

[0019] In one implementation scenario, as a possible approach, feature extraction can be performed directly based on the sensor data at the first moment to obtain the feature manifold at that moment. For example, neural network models such as convolutional neural networks and long short-term memory networks can be used to extract features from the sensor data at the first moment to obtain the feature manifold at that moment.

[0020] In another implementation scenario, as another possible approach, the sensing signals from several sensors in the sensor data may exhibit time discrepancies. For example, the sampling frequency of an inertial measurement unit (IMU) is above 200 Hz, while the output frequency of a visual recognition system such as a camera is 30 Hz. During vehicle operation (e.g., at a speed of 60 km / h), this time discrepancy can lead to significant distance errors. To address this, time delay compensation processing can be introduced to synchronize the sensing signals of each sensor. Afterward, feature extraction can be performed on the sensing signals of each sensor at the same time to obtain the feature manifold for that specific moment. For example, the local time when the sensor generates the sensing signal can be determined based on the sensor's local clock. The corrected time of the sensor can be obtained based on the difference between the time the sensor receives the synchronization message from the central computing platform and the time the synchronization message is sent. Based on this, the local time when the sensor generates the sensing signal can be corrected using the corrected time, resulting in the global time when the sensor generates the sensing signal. Furthermore, signal alignment and signal completion can be performed based on the global times when each sensor generates its sensing signal, yielding the sensor data at the same moment. It should be noted that sensor data from each sensor at the same moment can be selected sequentially as the sensor data for the first moment. The above method uses the time difference between the sending time of the synchronization message from the central computing platform and the receiving time of the sensor's synchronization message as a correction time to correct the local time when the sensor generates the sensing signal, thus obtaining the global time. Based on this, signal alignment and signal completion are achieved, ensuring that, as far as possible, each sensor corresponds to the same driving state at the same point in time.

[0021] In a specific implementation scenario, for ease of description, the local time when the sensor generates the sensing signal can be denoted as T. local The corrected global time T global It can be represented as: (1) In the above formula (1), This represents the phase difference between the receiving and transmitting times. Indicates the physical delay on the line. This represents the correction time. For any sensed signal from any sensor, at the calculated global time T... global After that, you can add a global time T. global Timestamp marker.

[0022] In a specific implementation scenario, for different sensors, the global time T is determined separately based on their respective sensing signals. global Then, based on the global time T globalPerform signal alignment. For simplicity, consider an example involving sensor A and sensor B. The sensing signal of sensor A includes: sensing signal A1 (its global time T...). global 06:20:01, representing hours, minutes, and seconds respectively (the same applies below), and sensing signal A2 (its global time T). global (06:20:05), sensing signal A3 (its global time T) global (06:20:09), sensing signal A4 (its global time T) global The time is 06:20:13), and the sensing signal of sensor B includes: sensing signal B1 (its global time T). global (06:20:01), sensing signal B2 (its global time T) global (06:20:09). Then, based on the global time T of sensor A and sensor B... global The two sensing signals can be aligned. For example, sensing signal A1 can be aligned with sensing signal B1 (meaning they correspond to the same driving state at the same time point 06:20:01), and sensing signal A3 can be aligned with sensing signal B2 (meaning they correspond to the same driving state at the same time point 06:20:09). Of course, the above examples are only one possible example of signal alignment in practical applications. Other possible situations are not limited here, nor will they be listed one by one.

[0023] In a specific implementation scenario, signal completion can be performed after signal alignment. The sensor with the highest sampling frequency can be used as the reference sensor, and all other sensors can be used as sensors to be completed. The target time for signal completion between adjacent times of the sensors to be completed can be determined by the global time of the actual sensing signal of the reference sensor between adjacent times. Furthermore, signal completion can be performed on the target time between adjacent times of the sensors to be completed based on cubic spline interpolation. For ease of understanding, we will still use the sensing signals of sensor A and sensor B as examples to illustrate signal completion. In this example, since sensor A has a higher sampling frequency than sensor B, sensor A can be used as the reference sensor and sensor B as the sensor to be completed. The target time for signal completion between adjacent times 06:20:01 and 06:20:09 of sensor B can be determined by the global time 06:20:05 of the actual sensing signal (i.e., sensing signal A2) of the reference sensor between the adjacent times 06:20:01 and 06:20:09. That is, the target time for signal completion between adjacent times 06:20:01 and 06:20:09 of sensor B is 06:20:05. At this time, signal completion can be performed on the target time 06:20:05 between adjacent times 06:20:01 and 06:20:09 of sensor B. Of course, the above example is merely one possible illustration of signal completion in practical applications. Other possible scenarios are not limited here, nor will they be listed individually. It should be noted that when performing signal completion based on cubic spline interpolation, it is possible to perform signal completion within each small time interval [t]. i ,t i+1 Within [the specified area], construct a third-order polynomial: (2) In the above formula (2), This can be determined using known points. At the connection points of adjacent intervals, the first and second derivatives must be equal (i.e., the vehicle's speed and acceleration will not change abruptly). For specific details, please refer to the technical details of cubic spline sampling interpolation, which will not be elaborated here. Furthermore, when performing signal completion, features can first be extracted from the sensing signals of adjacent time points of the sensor to be completed. Then, signal completion is performed on the target time point between the extracted features from the sensing signals of adjacent time points to complete the signal features of the signal to be completed between the sensing signals of adjacent time points (i.e., not directly completing the sensing signal, but completing the signal features), so as to facilitate the subsequent formation of a feature manifold. In real-world road tests, the synchronized data stream can effectively eliminate sensor ghosting, and the positioning drift is reduced by approximately 15 centimeters at a vehicle speed of 60 kilometers per hour. Cubic spline interpolation can make the rolling force feature field smoother, preserving the physical frequency of suspension bounce when filling in missing frames. Furthermore, similar signal alignment and signal completion can be performed on the sample training data in subsequent model training to provide a higher quality sample library for subsequent parameter optimization algorithms, minimize spurious noise caused by sampling intervals, and directly improve the convergence efficiency of parameter optimization.

[0024] In another implementation scenario, as yet another possible approach, to minimize the original noise in the sensing signal before feature extraction (e.g., engine vibration and road surface unevenness causing clutter to mix into the sensor signal during vehicle operation), signal decomposition can be performed on the sensing signal to obtain several signal components and their center frequencies. Then, based on the correlation coefficient between the signal components and the sensing signal, it is determined whether to retain a signal component as a candidate component. Furthermore, based on whether the center frequency of the candidate component is within the dynamic range, it is determined whether to retain it as a target component. Finally, feature extraction can be performed based on the data points at the first time step to obtain the feature manifold at that time step. It should be noted that the data points at the first time step can include the sensor's target component. This method, by decomposing the sensing signal and calculating the correlation coefficient between the signal components and the sensing signal to determine whether to retain a signal component as a candidate component, and then combining this with the determination of whether the center frequency of the candidate component is within the dynamic range to determine whether to retain it as a target component, allows for feature extraction using data points containing the target component to obtain the state manifold. This achieves component filtering from both the correlation coefficient and dynamic range dimensions, thus ensuring data purity through a double-layered protection.

[0025] In a specific implementation scenario, signal decomposition can be achieved using techniques such as VMD (Variational Mode Decomposition). Taking signal decomposition using VMD as an example, its essence is to solve a variational constraint problem, decomposing the original signal (in this embodiment, the sensor's sensing signal) into several sub-signals (i.e., signal components) with specific center frequencies. These sub-signals are called Intrinsic Mode Functions (IMFs). The bandwidth of each sub-signal is as narrow as possible, meaning that each component must be distributed around a center frequency. For ease of description, the above process can be represented by the following mathematical model: (3) In the above formula (3), This represents the k-th signal component. Represents signal components The center frequency, Represents the Dirac function, The above formula represents the convolution operation, where K represents the total number of signal components. The purpose of this formula is to obtain the single-sided spectrum of the sensed signal, then shift the spectrum to baseband using exponential correction, and finally calculate the square norm of the gradient to estimate the bandwidth. In this process, the Alternating Direction Method of Multipliers (ADMM) can be used to find the answer (i.e., the signal components and their center frequencies), which involves continuously updating each signal component. and its center frequency Through repeated testing, several smooth curves that best represent the original signal are found. In each update, the components can be adjusted based on the current residual. Compared to existing related technologies that are prone to mode aliasing problems, the above method can separate high-frequency mechanical noise from low-frequency vehicle trends as much as possible. It should be noted that, as a possible example, in the various signal components obtained through signal decomposition, the low-frequency components usually represent the actual dynamic trend of the vehicle, while the high-frequency components are often caused by sensor electronic noise or high-frequency engine vibration. Therefore, high-frequency signals can be filtered out, retaining the useful components that can reflect the rolling force characteristics.

[0026] In a specific implementation scenario, after obtaining the signal components and their center frequencies through signal decomposition, the correlation coefficient between the signal components and the sensed signal can be calculated to determine whether to retain the signal component as a candidate component. For example, the correlation coefficient r... k It can be represented as: (4) In the above formula (4), This indicates that the k-th signal component is at the i-th frequency point. The signal value, This represents the signal mean of the k-th signal component. This represents the average frequency of the k-th signal component. It should be noted that the correlation coefficient r... k The closer a value is to 1, the stronger the signal component. The better it represents the main trend of the original signal (i.e., the sensed signal). Of course, the above example is just one possible example of calculating the correlation coefficient in practical applications. Other methods for calculating the correlation coefficient are not limited here, nor will they be listed one by one.

[0027] In a specific implementation scenario, after calculating the correlation coefficient between each signal component and the sensed signal, it can be determined whether to retain a signal component based on whether the correlation coefficient is higher than a correlation threshold (e.g., 0.4). For example, if the correlation coefficient is higher than the threshold, the signal component can be retained; conversely, if the correlation coefficient is lower than the threshold, the signal component can be discarded. Taking accelerometers as an example, in the empty and fully loaded tests of autonomous trucks, the accelerometers generate severe vibrations when the truck passes over speed bumps. These vibrations, decomposed into higher-order components, have a correlation coefficient of only about 0.1 with the accelerometer's sensed signal, so these higher-order components can be discarded. Of course, the above example is merely one possible example of screening candidate components in practical applications; other possible methods are not limited here, nor will they be listed in detail.

[0028] In a specific implementation scenario, the dynamic range can characterize the frequency range of the effective signal. As mentioned earlier, higher-order components are usually caused by factors other than normal driving, such as vehicle vibration, while the change in rolling force is usually very slow. Therefore, the dynamic range can be set to 0 to 10 Hz. That is, when the center frequency of a candidate component exceeds the dynamic range, the candidate component can be eliminated; conversely, when the center frequency of a candidate component is within the dynamic range, the candidate component can be retained as the target component. Of course, the above example is only one possible example of screening target components in practical applications. Other possible implementation methods are not limited here, nor will they be listed one by one.

[0029] In a specific implementation scenario, after obtaining the data points at the first moment based on the target components in the sensing signals of each sensor, features can be extracted from these data points to obtain the feature manifold at the first moment. Specifically, an objective function can be constructed based on the difference between the weighted sum of the corresponding data points at each second moment and the data points at the first moment, with the weighted coefficients serving as unknowns in the objective function. The weighted coefficients at each second moment can then be solved based on the objective function to obtain the weighted coefficients at each second moment, and the feature vectors mapped to the low-dimensional space of the data points at each second moment. This yields the feature vectors mapped to the low-dimensional space of the data points at the first moment. These feature vectors can be obtained during the rolling force prediction process at the second moment. Furthermore, phase space reconstruction can be performed on the sub-feature vectors belonging to each sensor within the feature vectors at the first moment to obtain the phase space trajectories of each sensor at the first moment. Based on these phase space trajectories, the feature manifold at the first moment can be obtained. The above method first constructs an objective function with weighting coefficients as unknowns, and then obtains the feature vector of the data points at the first moment mapped to a low-dimensional space based on this function. The feature manifold at the first moment is then constructed through the phase space. Compared with related technologies that only focus on the current sensor readings during feature extraction, this method can take into account the synergistic effect and time difference between sensors. Combining the feature vector of the data points at the second moment mapped to a low-dimensional space with the phase space reconstruction can also reflect the evolution trend of the sensor readings, which helps to transform the messy raw sensor readings into high-quality dynamic features.

[0030] For example, for ease of description, the data point at the first moment can be denoted as X. i The weighting coefficient at the j-th second time step can be denoted as W. ij The j-th data point at the second time step can be X. j Then the objective function It can be represented as: (5) In formula (5) above, the sum of all weighting coefficients can equal 1. It should be noted that this set of weighting coefficients can characterize the geometric features of local data and is an inherent property of the data manifold. It should also be noted that minimizing the objective function minimizes the reconstruction error, thus obtaining the weighting coefficients W. ij .

[0031] For example, after obtaining the weighting coefficients, the feature vectors mapping the data points at the first moment to a low-dimensional space (e.g., 2D, 3D, etc.) can be obtained by following the same geometric logic as the aforementioned objective function. Referring to formula (5), the expression following the same geometric logic can be expressed as: (6) In the above formula (6), Y i Y represents the feature vector of the data point Xi at the first time step mapped to the low-dimensional space. j X represents the data point X at the j-th second time step. j The feature vectors are mapped to a low-dimensional space. After constructing an expression such as Equation (6), the data point X can be obtained by solving the eigenvalue problem. i Feature vectors mapped to a low-dimensional space. It should be noted that feature vectors mapped to a low-dimensional space can be regarded as key features after removing redundant information from the original data, used to capture the feature changes of vehicles under different road conditions.

[0032] For example, after extracting the feature vector mapped to the low-dimensional space for the data points at the first moment, the sub-feature vectors belonging to each sensor can be extracted from this vector. Then, the phase space can be reconstructed to obtain the corresponding phase space trajectory. For ease of description, the phase space trajectory X(n) of the sensor can be represented as: (7) In the above formula, x(n) represents the sampled value at the first moment (i.e., the sub-feature vector), τ represents the delay time, and m represents the embedding dimension. It should be noted that the phase space trajectory can not only describe the current state of the vehicle but also record its evolution trajectory over a past period. This processing method transforms static data points into dynamic trajectory lines. Furthermore, after obtaining the phase space trajectories of each sensor at the first moment, these trajectories can be combined to form the feature manifold for the first moment. To determine the aforementioned delay time τ, the mutual information method can be used to find a time point where past and present data have some correlation but are not completely repetitive, helping to ensure the model acquires the maximum amount of information. In addition, the spurious nearest neighbor method can be used to determine the embedding dimension m. If the dimension is too low, the trajectory lines may overlap; if the dimension is too high, the computational cost becomes enormous. The spurious nearest neighbor method can automatically find a critical point where the data trajectory fully unfolds in space, revealing the intrinsic laws governing rolling force changes. Taking a long downhill braking test as an example, the tire temperature will continuously rise. As the rubber softens, the rolling resistance will change accordingly, and relying solely on current sensor readings cannot accurately capture this trend. It should be noted that the specific process of phase space reconstruction can be found in the technical details of the mutual information method and the spurious nearest neighbor method, which will not be elaborated here. Furthermore, under standard test conditions, the signal-to-noise ratio is improved by approximately 30% after time delay compensation and VMD processing, and the construction of the feature manifold reduces the input dimension by 50% while retaining more than 95% of the original feature information. These improvements directly reduce the burden on the computing platform, which is crucial for autonomous driving embedded systems with limited computing power. Through the above series of operations, the chaotic raw sensor readings can be transformed into high-quality dynamic features.

[0033] Step S12: Analyze the characteristic manifold at the first time step and the characteristic manifold at the second time step before the first time step to obtain the sliding variance of the characteristic manifold within the target sliding window.

[0034] In this embodiment, the target sliding window can cover the first moment and the second moment preceding it. The rolling force prediction model can include a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, while the Gaussian kernel is responsible for capturing local instantaneous jumps (e.g., minor resistance fluctuations caused by road surface potholes). The first weight and the second weight can be negatively correlated. That is, the rolling force prediction model can adopt an HKSVR (Hybrid Kernel Support Vector Regression) structure. Furthermore, as mentioned earlier, in practical applications, each moment when the sensor acquires data can be selected sequentially as the first moment to obtain the predicted rolling force at each moment. When predicting the rolling force at the second moment before the first moment, the characteristic manifold of the second moment can be obtained. The method of obtaining this manifold can be found in the aforementioned description of the characteristic manifold of the first moment, which will not be repeated here.

[0035] In one implementation scenario, for ease of description, the hybrid kernel function K can be... mix (x i ,x j ) is represented as: (8) In the above formula (8), K poly Denotes the polynomial kernel, K RBF Let x represent the Gaussian kernel, xi represent the characteristic manifold at time i (e.g., time 1), and x j Let K represent the characteristic manifold or the j-th support vector at time j (e.g., time 2). λ represents the first weight, and 1-λ represents the second weight. These two weights determine the contribution ratio of the polynomial kernel and the Gaussian kernel. During rolling force prediction, the sliding variance of the characteristic manifold can be calculated in real time to determine the specific values ​​of the first and second weights. The polynomial kernel K... poly It can be represented as: (9) In addition, Gaussian kernel K RBF It can be represented as: (10) It should be noted that the above descriptions are merely illustrative of polynomial kernels and Gaussian kernels in mixed kernel functions. For the specific concepts of polynomial kernels and Gaussian kernels, please refer to the technical details of polynomial kernels and Gaussian kernels, which will not be repeated here.

[0036] In one implementation scenario, to ensure the smooth and continuous evolution of the aforementioned weighting factors across different operating conditions, a target sliding window can be pre-set. This target sliding window covers the first time step and the second time step preceding it. The specific length of the target sliding window is not limited. The characteristic manifolds of the first time step and the second time step preceding it within the target sliding window can then be analyzed to obtain the sliding variance of the characteristic manifolds within the target sliding window. For example, the characteristic manifolds of the first time step and the second time step preceding it can be arranged chronologically. Then, the variance of the characteristic manifolds of the first time step and the second time step preceding it within the target sliding window can be calculated to obtain the sliding variance of the characteristic manifolds within the target sliding window. It should be noted that the target sliding window can slide over time; therefore, the length and sliding step of the target sliding window are not limited.

[0037] Step S13: Based on the sliding variance, determine the first value of the first weight at the first time, and based on the first value and the negative correlation, determine the second value of the second weight at the first time.

[0038] In this embodiment of the disclosure, the sliding variance can be negatively correlated with the first weight. Thus, the first value of the first weight at the first time can be calculated based on this negative correlation. Then, the second value of the second weight at the first time can be calculated by combining the negative correlation between the first weight and the second weight (e.g., the sum of the first weight and the second weight is 1).

[0039] In one implementation scenario, as a possible approach, to determine the first value of the first weight and the second value of the second weight, the variance difference between the sliding variance and the variance threshold can be obtained first. Then, the reciprocal of the sum of the weight base (based on the natural constant and exponentially by the variance difference) and 1 can be obtained as the first value of the first weight at the first time step. Furthermore, considering the negative correlation between the first and second weights, the second value of the second weight at the first time step can be calculated based on this first value (e.g., subtracting the first value of the first weight at the first time step from 1 yields the second value). For ease of description, the first weight λ is represented as: (11) In the above formula (11), Represents the sliding variance. Indicates the variance threshold. This represents the weight base. The above method obtains the variance difference between the sliding variance and the variance threshold, and then obtains the weight base (based on the natural constant and exponentially by the variance difference) and the reciprocal of the sum of these two values. This is used as the first value of the first weight at the first moment, enabling timely detection of changes when the sliding variance exceeds the variance threshold and prompting a reduction in the first value of the first weight.

[0040] In another implementation scenario, as another possible implementation method, in order to determine the first value of the first weight and the second value of the second weight, the variance difference between the sliding variance and the variance threshold can also be obtained. Based on this, the first weight base with the natural constant as the base and the variance difference as the exponent can be obtained, and the second weight base with the natural constant as the base and the negative number of the variance difference as the exponent can be obtained. Thus, the ratio between the difference between the first weight base and the second weight, and the sum of the first weight base and the second weight base can be obtained as the first value of the first weight at the first time. Then, combined with the negative correlation between the first weight and the second weight, the second value of the second weight at the first time can be calculated based on the first value of the first weight at the first time (e.g., 1 can be subtracted from the first value of the first weight at the first time to obtain the second value of the second weight at the first time).

[0041] It should be noted that the above examples are merely a few possible examples of calculating the values ​​of the first and second weights in practical applications. Other possible calculation methods are not limited here, nor will they be listed in detail. Furthermore, after calculating the first value of the first weight and the second value of the second weight at the first time step, these two values ​​can be substituted into the hybrid kernel function to obtain the rolling force prediction model for the first time step. As one possible example, the update frequency of the first and second weights can be synchronized with the clock of the vehicle chassis control domain. Of course, the above examples are only one possible example of the update frequency of the first and second weights in practical applications; other possible update frequencies are not limited here, nor will they be listed in the same way.

[0042] Step S14: Based on the rolling force prediction model, predict the characteristic manifold at the first moment to obtain the predicted rolling force at the first moment.

[0043] In one implementation scenario, after substituting the first value of the first weight and the second value of the second weight at the first time step into the aforementioned hybrid kernel function, the rolling force prediction model for the first time step can be obtained. At this point, the feature manifold for the first time step can be predicted based on the rolling force prediction model to obtain the predicted rolling force. For ease of understanding, suppose an autonomous truck is traversing a long uphill road. Due to the relatively fixed slope, the global component of the rolling force dominates. In this case, the multinomial kernel plays a major role, ensuring the prediction curve is as smooth and stable as possible. Suddenly, a waterlogged area appears on the road, causing an instantaneous change in tire adhesion characteristics. The slight slippage from the wheel speed sensor is immediately reflected in the feature manifold. The adaptive factor detects this surge in local features and rapidly increases the weight ratio of the Gaussian kernel (i.e., the second value of the second weight becomes relatively larger) to allow the rolling force prediction model to quickly capture this abrupt change in resistance. Furthermore, it should be noted that this hybrid kernel function also brings an additional benefit: it effectively solves the overfitting problem during model training. Using a single Gaussian kernel function as a comparison can easily lead to overfitting to noise in the training set if the parameters are not set properly. The addition of a multinomial kernel provides a structured constraint, forcing the model to adhere to global physical logic while focusing on details. This mutually restraining mechanism improves the model's generalization performance. Comparative tests on different types of road surfaces show that in complex scenarios such as muddy and gravel roads, the prediction error is reduced by approximately 18% compared to traditional models. Furthermore, before using the rolling force prediction model to predict the feature manifold at the first moment, the rolling force prediction model can be trained first; details are described below and will not be elaborated here.

[0044] In one implementation scenario, when the current time reaches the third time after the first time, the third time can be regarded as the new first time. At this time, we can return to the steps of extracting features based on the sensor data of the first time to obtain the feature manifold of the first time, so as to update the first value of the first weight and the second value of the second weight in the new first time. After substituting into the aforementioned hybrid kernel function, we can obtain the rolling force prediction model of the new first time. At this time, we can predict the feature manifold of the new first time based on the rolling force prediction model to obtain the predicted rolling force of the new first time. By repeating this process, we can obtain the predicted rolling force in real time.

[0045] In one implementation scenario, during real-time prediction, the difference between the predicted rolling force and the actual rolling force at the first moment can be obtained. If the rolling force difference exceeds the tolerance range, the feature manifold at the first moment can be selected as the new support vector for the rolling force prediction model. Based on the feature distances between the new support vector and each of the old support vectors in the rolling force prediction model, an association vector is obtained. This association vector can then be used to update the sensitivity matrix of the rolling force prediction model, resulting in a new sensitivity matrix. This sensitivity matrix reflects the rolling force prediction model's sensitivity to data changes. Furthermore, based on the new sensitivity matrix and the rolling force difference, the adjustment range of the weight parameters in the rolling force prediction model can be determined. This method, when the rolling force difference exceeds the tolerance range, updates the sensitivity matrix of the rolling force prediction model using the feature manifold at the first moment, thereby incrementally updating the parameters in the rolling force prediction model. This allows the rolling force prediction model to detect new samples that break the current prediction rules during driving, enabling self-updating. Since incremental updates are used without full training, it helps to ensure that the computing power of the onboard chip is not overloaded.

[0046] In a specific implementation scenario, for ease of description, the predicted rolling force of the rolling force prediction model for the characteristic manifold at the first moment can be denoted as f(x). c Then, the constraint condition for incremental updates during real-time prediction can be expressed as: (12) In the above formula (12), y c This represents the actual rolling force measured by the sensor. This indicates the fault tolerance range. If the rolling force difference is within the fault tolerance range, the sample (i.e., the feature manifold at the first moment) can be discarded so that the onboard chip does not waste computing power to process small noise. Conversely, if the rolling force difference exceeds the fault tolerance range, the sample (i.e., the feature manifold at the first moment) can be absorbed as a new support vector.

[0047] In a specific implementation scenario, after selecting the feature manifold at the first time step as the new support vector, the new support vector and any old support vector can be used as input parameters to the mixing kernel function to obtain the feature distance between the new support vector and the corresponding old support vector. Then, based on the feature distances corresponding to each old support vector, the association vector is obtained. For example, the feature distances corresponding to each old support vector can be used as the feature elements in the association vector. For ease of description, the association vector can be denoted as β.

[0048] In a specific implementation scenario, after obtaining the correlation vector, the sensitivity matrix can be updated to obtain a new sensitivity matrix for the rolling force prediction model. For example, a correction matrix for updating the sensitivity matrix can be obtained based on the correlation vector and its transpose. Then, the sensitivity matrix is ​​updated based on the product of the forgetting factor and the correction matrix to obtain the new sensitivity matrix R. new : (13) In the above formula (13), R represents the sensitivity matrix, β represents the correlation vector, and β T β represents the transpose of the associated vector. T Let represent the correction matrix, and 1 / γ represent the forgetting factor. Furthermore, after obtaining the new sensitivity matrix, the adjustment magnitude Δθ of the weight parameters in the rolling force prediction model can be obtained by combining it with the rolling force difference. (14) In the above formula (14), R new Let represent the new sensitivity matrix, and 'e' represent the rolling force difference. The above process allows the model to complete weight adjustments within milliseconds, transforming complex matrix operations into simple vector additions, thus reducing the computational cost from O(N) to O(N). 3 ) reduced to O(N 2 This efficiency improvement is crucial for embedded controllers. Furthermore, in practical applications, a fixed-length sliding window can be used to control the model size, retaining only core samples within the most recent distance. As new road condition data enters the model, older data is gradually removed. This is because changes in rolling force have local correlations; for example, road surface features from five minutes ago have little reference value for the current situation. This metabolic mechanism can prevent the model from expanding indefinitely, maintaining a streamlined and efficient operating state. It should be noted that the sensitivity matrix (e.g., the covariance matrix) can be manually set at the initial moment.

[0049] The above scheme extracts features from sensor data at the first moment to obtain the feature manifold at that moment. Then, it analyzes the feature manifold at the first moment and the feature manifold at the second moment prior to the first moment to obtain the sliding variance of the feature manifold within the target sliding window. The target sliding window covers the first moment and the second moment prior to the first moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of rolling force, while the Gaussian kernel is responsible for capturing local instantaneous jumps. The first and second weights are negatively correlated. Therefore, based on the sliding variance, the first value of the first weight at the first moment is determined, and based on the first value and the negative correlation, the second weight at the first moment is determined. The second value at time step 1 is used, and the sliding variance is negatively correlated with the first weight. Then, based on the rolling force prediction model, the characteristic manifold at time step 1 is predicted to obtain the predicted rolling force at time step 1. On the one hand, the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with the first and second weights respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps, which helps to maximize adaptability to complex working conditions. Furthermore, since only the characteristic manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, it also helps to reduce the computational complexity of rolling force prediction. On the other hand, the characteristic manifold extracted from sensor data is adaptively... The first value of the first weight and the second value of the second weight should be determined accordingly. Since the sliding variance of the characteristic manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can be determined which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the data feedback from the sensors fluctuates less, and the local variability of the characteristic manifold is low, resulting in a smaller sliding variance. Therefore, the first value of the first weight can be adaptively increased while the second value of the second weight can be decreased, allowing the rolling force prediction model to focus more on the polynomial kernel to accurately control the long-term trend. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensors fluctuates at high frequencies, and the special... As the eigenvariance increases rapidly, the adaptive weights react immediately. The first value of the first weight decreases adaptively, and the second value of the second weight increases adaptively. This allows the rolling force prediction model to switch to a model that focuses more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted to follow the actual working conditions, which helps improve the accuracy of rolling force prediction. On the other hand, when calculating the eigenvariance, the sliding variance is calculated within the target sliding window by combining the eigenmanifold at the first moment and the eigenmanifold at the second moment before the first moment. This can filter out instantaneous sampling noise and help ensure that the evolution of the weight factors is as smooth and continuous as possible between different working conditions, which also helps improve the accuracy of rolling force prediction.Therefore, it can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction.

[0050] In some disclosed embodiments, such as those described above, before using the rolling force prediction model to predict the characteristic manifold at the first moment, the rolling force prediction model can be trained first. For example, the target hyperparameters for model training can be obtained by searching the hyperparameter solution space of the rolling force prediction model. Then, based on the sample training data, the rolling force prediction model is trained and optimized under the control of the target hyperparameters. This allows the absolute value of the difference between the predicted rolling force of the rolling force of the rolling force prediction model on the same sample training data before and after the current round of training and optimization to be obtained as the stable value of the current round of training and optimization. Then, based on the average value of the stable values ​​of several rounds of training and optimization, it can be determined whether to perform Cauchy mutation on the target hyperparameters to obtain new target hyperparameters, so that the training and optimization can be re-executed under the control of the new target hyperparameters. The above method obtains the target hyperparameters by searching the hyperparameter solution space, and then trains and optimizes the rolling force prediction model accordingly. This allows us to obtain the absolute value of the difference between the predicted rolling force of the rolling force prediction model and the same sample training data before and after the current training round, and to determine whether to perform Cauchy mutation on the target hyperparameters to obtain new target hyperparameters, what mechanism can increase the diversity of the solution space, and to ensure that the optimization process is not interrupted by a local disturbance as much as possible.

[0051] In one implementation scenario, the target hyperparameters can include, but are not limited to, penalty factors, insensitive loss coefficients, and weight ratios of the hybrid kernel function. In other words, any set of hyperparameter solutions can be considered as a particle in the hyperparameter solution space. As one possible implementation, the target hyperparameters can be obtained by collaboratively searching using Particle Swarm Optimization (PSO) and Beetle Antennae Search (BAS). Specifically, the target hyperparameters can be obtained by searching the hyperparameter solution space using PSO-BAS. For example, the search can first be performed using PSO until the particle swarm's movement speed falls below a threshold. Then, the optimal solution found by the particle swarm can be used as the starting point for high-precision sampling by the Beetle Antennae Search to ultimately obtain the target hyperparameters. It's important to note that the collaborative search between the Particle Swarm Optimization (PSO) algorithm and the beetle whisker algorithm is not a simple sequential process, but rather a collaborative mechanism based on spatial scale switching. During the PSO execution phase, the PSO algorithm primarily handles a macroscopic global scan, treating each set of hyperparameters as a particle in space. A large number of particles are randomly distributed in the parameter space, each with its own position and velocity. As the particles fly, they continuously evaluate their current fitness value, using the mean square error of the rolling force prediction as an evaluation criterion. The particles remember their historical best positions and also receive position information from the best particles in the swarm. This mechanism simulates the collaborative behavior of swarm intelligence, rapidly narrowing the search range and eliminating parameter regions with extremely poor performance. Furthermore, the beetle whisker algorithm is a biomimetic algorithm that mimics the behavior of longhorn beetles sensing the concentration of food odors through their antennae. It determines its next direction by comparing the signal strength received by its left and right antennae. It can adaptively adjust its search step size to approach the optimal solution. If the detected fitness value improves, the beetle can accelerate its movement; if the fitness value deteriorates, it can quickly shrink the search radius. It exhibits extremely high search efficiency and sensitivity in local spaces. Through the collaborative work of particle swarm optimization and the beetle whisker algorithm, convergence can be achieved in just a few hundred iterations. Compared to existing grid search algorithms, which require exhaustive searches of all combinations and consume hours or even days of computation time, this reduces computational overhead by more than 60% and improves prediction accuracy by 20%. It also avoids the trap of traditional algorithms easily getting stuck in local optima, ensuring optimal parameter states under various extreme conditions. Of course, the above example is merely one possible instance of searching the hyperparameter solution space to obtain the target hyperparameters in practical applications. Other possible search methods are not limited here, nor will they be listed in detail.

[0052] In a specific implementation scenario, as mentioned earlier, any set of hyperparameter solutions can be considered as a particle in the hyperparameter solution space. All particles can then move together within the parameter space, and particle swarm optimization can be responsible for finding the approximate region of the optimal solution. The position of each particle can be updated based on the following formula: (15) (16) In the above formulas (15) and (16), v represents velocity, x represents position, pbest represents the best position in the particle's history, and gbest represents the best position of the swarm. For the specific meanings of other parameters, please refer to the technical details of the particle swarm algorithm, which will not be repeated here. The particles can quickly identify invalid areas in the above way. They will lock in the parameter zone with the smallest rolling force prediction error. The particle swarm will slow down when it gets close to the target. When the particle swarm's movement speed is lower than the threshold, the switching procedure can be triggered, and the optimal solution found by the particle swarm can be used as the starting point of the beetle whisker algorithm.

[0053] In a specific implementation scenario, as mentioned earlier, the longhorn beetle whisker algorithm is a biomimetic algorithm. Longhorn beetles do not rely on complex maps when searching for food; instead, they use their two antennae on their heads to sense odor intensity. If the left antennae detects a stronger odor, they turn left; otherwise, they turn right. By simulating this behavior, rapid detection can be achieved within a very small area. For example, a random search direction can be generated first, which can be achieved using a high-dimensional unit vector. express: (17) In the above formula (17), the rnd function is used to generate a set of random numbers, dim represents the total number of elements in this set of random numbers, ||.|| represents the modulus of the vector, and the unit vector can represent the longhorn beetle randomly rotating its head to determine the baseline for detection, based on which the spatial positions of the left and right tentacles can be determined: (18) (19) In the above formulas (18) and (19), x t This indicates the current position, and d0 represents the distance between the two tentacles. Thus, we can determine the positions at x and d0 respectively. L x R The prediction error between the two locations is similar to a longhorn beetle comparing the odor concentration on its left and right sides, and adjusting its next direction based on the sensing results: (20) In the above formula (20), f(x) represents the prediction error function, and sign represents the sign function. If the error on the left is small, the beetle moves to the left; otherwise, it moves to the right. Furthermore, This indicates the step size. As the number of searches increases, the step size can be continuously reduced. The numerical value of this design ensures that the beetle eventually stops stably at the optimal solution. Of course, the above description is merely an illustrative example of the beetle whisker algorithm; for details, please refer to the technical specifications of the beetle whisker algorithm, which will not be elaborated upon here.

[0054] In one implementation scenario, after obtaining the target hyperparameters for training the rolling force prediction model, the model can be trained and optimized based on sample training data under the control of the target hyperparameters. It should be noted that the sample training data can include sensor data at each sample time point, and the sensor data at each sample time point can be labeled with the actual rolling force. The training loss of the rolling force prediction model can be obtained by measuring the difference between the predicted rolling force and the actual rolling force. Thus, during the training and optimization process, the model parameters of the rolling force prediction model can be adjusted based on the training loss.

[0055] In one implementation scenario, after training and optimizing the rolling force prediction model based on sample training data according to the target hyperparameters, the absolute value of the difference between the predicted rolling force of the rolling force prediction model before and after the current round of training and optimization on the same sample training data can be obtained as the stable value of the current round of training and optimization. Based on the average of the stable values ​​of several rounds of training and optimization, it is determined whether to perform Cauchy mutation on the target hyperparameters to obtain new target hyperparameters, so that training and optimization can be re-executed under the control of the new target hyperparameters. For example, if the average of the stable values ​​of several rounds of training and optimization is lower than a preset threshold, it can be determined that Cauchy mutation should be performed on the target hyperparameters to obtain new target hyperparameters. Conversely, if the average of the stable values ​​of several rounds of training and optimization is not lower than the preset threshold, Cauchy mutation should not be performed on the target hyperparameters, so that the rolling force prediction model can continue to be trained and optimized under the control of the original target hyperparameters until the rolling force prediction model training converges.

[0056] In a specific implementation scenario, the possibility that the aforementioned optimization algorithm may enter an idling state in the later stages of practical application cannot be ruled out. For example, a longhorn beetle may linger in a certain trough but cannot see the deeper pit (i.e., a better solution) further away. In order to ensure that the model can escape the trap as much as possible, a dynamic feedback layer can be introduced, which can monitor the changes in the predicted rolling force in real time. For example, a sliding window can be used to record the fitness values ​​of the most recent M iterations and calculated using the following formula: (twenty one) In the above formula (21), f(x) if(x) represents the predicted rolling force of the rolling force prediction model based on the sample training data in the i-th iteration. i-1 Let represent the predicted rolling force of the rolling force prediction model based on the sample training data at the (i-1)th iteration. The difference between the two can be used as the stable value at the i-th iteration. ΔF represents the average of the stable values ​​over M iterations. It should be noted that if ΔF is less than a preset threshold, the algorithm can be considered to have fallen into a local optimum. At this point, simple training optimization is meaningless, and Cauchy mutation needs to be performed on the current position (i.e., the target hyperparameter). This is equivalent to giving a huge push to a longhorn beetle stuck in a dead end.

[0057] In a specific implementation scenario, to perform Cauchy mutation on the target hyperparameters to obtain new target hyperparameters, a scaling factor can be obtained based on the product of the step size parameter and a random number conforming to the Cauchy distribution. The target hyperparameters are then scaled based on this scaling factor to obtain the new target hyperparameters. For ease of description, the new target hyperparameter x... new It can be represented as: (twenty two) In the above formula (22), x old Let C(0,1) represent the old target hyperparameters, C(0,1) represent random numbers conforming to the Cauchy distribution, and α represent the step size parameter. The above method obtains a scaling factor based on the product of the step size parameter and the random numbers conforming to the Cauchy distribution. Then, the target hyperparameters are scaled based on this scaling factor to obtain new target hyperparameters. Since the Cauchy distribution has long tails at both ends, it allows particles to potentially make significant jumps. These jumps allow the search point to escape current obstacles and enter a new parameter region, helping to increase the diversity of the solution space and ensuring that the optimization process is not interrupted by local disturbances.

[0058] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the driving control method of this application. It should be noted that the process operations in this embodiment can be executed by an electronic device with computing capabilities or related equipment containing an electronic device. The specific structure and type of the electronic device and related equipment containing an electronic device are not limited herein. Furthermore, this embodiment can be applied to application scenarios including but not limited to autonomous driving, and the application scenarios of this embodiment are not limited herein. Specifically, this embodiment may include the following steps: Step S21: Make a prediction based on the sensor data at the current moment to obtain the predicted rolling force at the current moment.

[0059] In this embodiment of the disclosure, the predicted rolling force at the current moment is obtained based on the above-described rolling force prediction method embodiment. For details, please refer to the aforementioned rolling force prediction method embodiment, which will not be repeated here.

[0060] Step S22: Based on the predicted rolling force at the current moment, perform torque control on the vehicle motor.

[0061] Specifically, the predicted rolling force can be directly fed into the vehicle's control layer. This means the predicted rolling force can be introduced as a feedforward term into the controller, allowing it to be anticipated before resistance actually affects vehicle speed. The control layer can adjust the motor's output torque in advance based on the predicted resistance magnitude. This proactive compensation mechanism significantly improves speed tracking accuracy and helps reduce the bumpy feeling when the vehicle is driving on rough roads. For ease of description, consider a heavy-duty transport scenario: when an autonomous truck is driving in the rain, the rolling resistance on the road increases due to water accumulation. Sensors detect minute errors in the driving force, and the online incremental module immediately captures this trend, rapidly adjusting the bias term of the hybrid kernel function. This causes the predicted rolling force from the rolling force prediction model to increase accordingly. Upon receiving this signal, the control layer slightly increases the power output. This process occurs within a very short time, so much so that passengers hardly perceive the speed fluctuation. It's important to note that this closed-loop feedback ensures optimal energy distribution as much as possible, helping to avoid frequent and ineffective acceleration and deceleration. In addition, this closed-loop system plays a crucial role in safety. In practical applications, it can continuously monitor the residual between the predicted rolling force and the observed rolling force (e.g., acquiring the predicted rolling force at time t and using it for torque control, and collecting the observed rolling force through relevant sensors to calculate the difference between the predicted and observed rolling forces at time t). If the residual exceeds a preset threshold, a safety alarm mechanism can be triggered (e.g., alerting to sensor malfunctions, severe tire deflation, etc.). This data-driven fault diagnosis function adds an extra layer of safety protection to autonomous driving. Furthermore, during motion, online incremental updates can be performed as described in the aforementioned embodiments to fine-tune the model parameters of the rolling force prediction model based on actual motion deviations. This cyclical closed-loop process allows the vehicle to become a self-evolving intelligent agent.

[0062] The above scheme predicts the rolling force based on sensor data at the current moment. This predicted rolling force is obtained using the rolling force prediction method described above. Based on this predicted rolling force, torque control is applied to the vehicle motor. Firstly, the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, while the Gaussian kernel captures local instantaneous changes. This helps to maximize adaptability to complex operating conditions. Furthermore, since only the feature manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, the computational complexity of rolling force prediction is reduced. Secondly, since the first value of the first weight and the second value of the second weight are adaptively determined based on the feature manifold extracted from the sensor data, and since the sliding variance of the feature manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can determine which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the sensor... When the feedback data fluctuates relatively little, the local variability of the feature manifold is low, resulting in a small sliding variance. Therefore, the first value of the first weight and the second value of the second weight can be adaptively increased and decreased. This allows the rolling force prediction model to focus more on the polynomial kernel to accurately grasp the long-term trend. However, when encountering special road conditions such as speed bumps or potholes, the sensor feedback data generates high-frequency fluctuations, and the feature variance increases rapidly. Therefore, the adaptive weights will react immediately, that is, the first value of the first weight will be adaptively decreased and the second value of the second weight will be adaptively increased. This allows the rolling force prediction model to switch to focusing more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted according to the actual working conditions, which helps to improve the accuracy of rolling force prediction. On the other hand, when calculating the feature variance, the feature manifold at the first moment and the feature manifold at the second moment before the first moment are combined to calculate the sliding variance within the target sliding window. This can filter instantaneous sampling noise and help to ensure that the evolution of the weight factors between different working conditions is as smooth and continuous as possible, which also helps to improve the accuracy of rolling force prediction. Therefore, it can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction, thereby helping to improve the response speed and accuracy of driving control.

[0063] Please see Figure 3 , Figure 3This is a schematic diagram of a framework of an embodiment of the rolling force prediction device of this application. The rolling force prediction device 30 includes: a feature extraction module 31, a manifold analysis module 32, a weight update module 33, and a rolling prediction module 34. The feature extraction module 31 is used to extract features based on sensor data at a first time moment to obtain the feature manifold at the first time moment; the manifold analysis module 32 is used to analyze the feature manifold at the first time moment and the feature manifold at a second time moment before the first time moment to obtain the sliding variance of the feature manifold within a target sliding window; wherein, the target sliding window covers the first time moment and the second time moment before the first time moment, and the rolling force prediction model includes a first weight and a second weight respectively. A hybrid kernel function weighted by a polynomial kernel and a Gaussian kernel is used. The polynomial kernel is responsible for the long-term trend of rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps. The first weight and the second weight are negatively correlated. The weight update module 33 is used to determine the first value of the first weight at the first time based on the sliding variance, and to determine the second value of the second weight at the first time based on the first value and the negative correlation. The sliding variance is negatively correlated with the first weight. The rolling prediction module 34 is used to predict the characteristic manifold at the first time based on the rolling force prediction model to obtain the predicted rolling force at the first time.

[0064] In the above scheme, the rolling force prediction device 30 performs feature extraction based on sensor data at a first moment to obtain the feature manifold at the first moment. Then, it analyzes the feature manifold at the first moment and the feature manifold at a second moment prior to the first moment to obtain the sliding variance of the feature manifold within a target sliding window. The target sliding window covers the first moment and the second moment prior to the first moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps. The first weight and the second weight are negatively correlated. Therefore, based on the sliding variance, the first value of the first weight at the first moment is determined, and based on the first value and the negative correlation, the rolling force prediction model is determined. The second weight is the second value at the first time step, and the sliding variance is negatively correlated with the first weight. Based on the rolling force prediction model, the characteristic manifold at the first time step is predicted to obtain the predicted rolling force at the first time step. On the one hand, since the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with the first and second weights respectively, and the polynomial kernel is responsible for the long-term trend of the rolling force while the Gaussian kernel is responsible for capturing local instantaneous jumps, it helps to maximize adaptability to complex working conditions. Furthermore, since only the characteristic manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, it also helps to reduce the computational complexity of rolling force prediction. On the other hand, since the characteristic manifold extracted from sensor data... The characteristic manifold adaptively determines the first value of the first weight and the second value of the second weight. Since the sliding variance of the characteristic manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can determine which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the data feedback from the sensors fluctuates less, resulting in lower local variability of the characteristic manifold and thus a smaller sliding variance. Therefore, it can adaptively increase the first value of the first weight and decrease the second value of the second weight, allowing the rolling force prediction model to focus more on the polynomial kernel to accurately grasp long-term trends. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensors fluctuates at high frequencies. As the characteristic variance increases rapidly, the adaptive weights react immediately. The first value of the first weight decreases adaptively, and the second value of the second weight increases adaptively. This allows the rolling force prediction model to switch to a model that focuses more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted to follow the actual working conditions, which helps to improve the accuracy of rolling force prediction. On the other hand, when calculating the characteristic variance, the sliding variance is calculated within the target sliding window by combining the characteristic manifold of the first time step and the characteristic manifold of the second time step before the first time step. This can filter out instantaneous sampling noise and help to ensure that the evolution of the weight factors between different working conditions is as smooth and continuous as possible, which also helps to improve the accuracy of rolling force prediction.Therefore, it can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction.

[0065] In some disclosed embodiments, the weight update module 33 includes a difference calculation submodule for obtaining the variance difference between the sliding variance and the variance threshold; the weight update module 33 includes a weight calculation submodule for obtaining the reciprocal of the sum of the weight base with the natural constant as the base and the variance difference as the exponent, and 1, as the first value of the first weight at the first time.

[0066] In some disclosed embodiments, the rolling force prediction device 30 includes a residual calculation module for obtaining the rolling force difference between the predicted rolling force and the actual rolling force at a first moment; the rolling force prediction device 30 includes an association calculation module for selecting a feature manifold at the first moment as a new support vector for the rolling force prediction model in response to the rolling force difference exceeding the fault tolerance range, and obtaining an association vector based on the feature distance between the new support vector and each old support vector of the rolling force prediction model; the rolling force prediction device 30 includes a matrix update module for updating the sensitivity matrix of the rolling force prediction model based on the association vector to obtain a new sensitivity matrix of the rolling force prediction model; wherein, the sensitivity matrix reflects the sensitivity of the rolling force prediction model to data changes; the rolling force prediction device 30 includes an amplitude determination module for obtaining the adjustment amplitude of the weight parameters in the rolling force prediction model based on the new sensitivity matrix and the rolling force difference.

[0067] In some disclosed embodiments, the association calculation module includes a distance calculation submodule, used to obtain the feature distance between the new support vector and the corresponding old support vector by using the new support vector and any old support vector as input parameters of the mixing kernel function; the association calculation module includes a vector calculation submodule, used to obtain the association vector based on the feature distances corresponding to each old support vector.

[0068] In some disclosed embodiments, the matrix update module includes a correction calculation submodule for obtaining a correction matrix for updating the sensitivity matrix based on the association vector and the transpose of the association vector; the matrix update module includes a matrix calculation submodule for updating the sensitivity matrix based on the product of the forgetting factor and the correction matrix to obtain a new sensitivity matrix.

[0069] In some disclosed embodiments, the rolling force prediction device 30 includes a parameter search module for searching the hyperparameter solution space of the rolling force prediction model to obtain target hyperparameters for model training; the rolling force prediction device 30 includes a training optimization module for training and optimizing the rolling force prediction model based on sample training data under the control of the target hyperparameters; the rolling force prediction device 30 includes a stability measurement module for obtaining the absolute value of the difference between the predicted rolling force of the rolling force prediction model on the same sample training data before and after performing the current round of training optimization, as the stable value of the current round of training optimization; and the rolling force prediction device 30 includes a parameter mutation module for determining whether to perform Cauchy mutation on the target hyperparameters to obtain new target hyperparameters based on the average value of the stable values ​​of several rounds of training optimization, so as to re-execute the training optimization under the control of the new target hyperparameters.

[0070] In some disclosed embodiments, the parameter mutation module includes a factor calculation submodule for obtaining a scaling factor based on the product of the step size parameter and a random number conforming to the Cauchy distribution; the parameter mutation module includes a parameter scaling submodule for scaling the target hyperparameter based on the scaling factor to obtain a new target hyperparameter.

[0071] In some disclosed embodiments, the target hyperparameter includes at least one of the following hyperparameters: a penalty factor, an insensitive loss coefficient, and a weighting ratio of the hybrid kernel function; and / or, the target hyperparameter is obtained by searching in the hyperparameter solution space based on PSO-BAS; and / or, if the average value of the stable value after several rounds of training optimization is lower than a preset threshold, it is determined that Cauchy mutation is performed on the target hyperparameter to obtain a new target hyperparameter; and / or, the sample training data includes sensor data at each sample time, the sensor data at each sample time is labeled with the actual rolling force, and the training loss of the rolling force prediction model is obtained by measuring the difference between the predicted rolling force and the actual rolling force.

[0072] In some disclosed embodiments, the sensor data includes sensing signals from several sensors. The feature extraction module 31 includes a signal decomposition submodule for decomposing the sensing signals to obtain several signal components and the center frequency of each signal component. The feature extraction module 31 includes a first determination submodule for determining whether to retain a signal component as a candidate component based on the correlation coefficient between the signal component and the sensing signal. The feature extraction module 31 includes a second determination submodule for determining whether to retain a candidate component as a target component based on whether the center frequency of the candidate component is within the dynamic range. The feature extraction module 31 includes a manifold extraction submodule for performing feature extraction based on the data points at a first time moment to obtain the feature manifold at the first time moment. The data points at the first time moment include the target components of the sensors.

[0073] In some disclosed embodiments, the manifold extraction submodule includes a function construction unit, used to construct an objective function based on the weighted sum of the data points at each second time step and the difference between the data points at the first time step and the weighted coefficients at each second time step; wherein the weighted coefficients are unknowns in the objective function; the manifold extraction submodule includes a low-dimensional mapping unit, used to obtain the feature vector of the data points at the first time step mapped to the low-dimensional space based on the weighted coefficients at each second time step obtained by solving the objective function and the feature vector of the data points at each second time step mapped to the low-dimensional space; wherein the feature vector of the data points at the second time step mapped to the low-dimensional space is obtained during the rolling force prediction process at the second time step; the manifold extraction submodule includes a spatial reconstruction unit, used to perform phase space reconstruction based on the sub-feature vectors belonging to each sensor in the feature vector at the first time step, to obtain the phase space trajectory of each sensor at the first time step; the manifold extraction submodule includes a manifold acquisition unit, used to obtain the feature manifold at the first time step based on the phase space trajectory of each sensor at the first time step.

[0074] In some disclosed embodiments, the sensor data includes sensing signals from several sensors. The rolling force prediction device 30 includes a time determination module, used to determine the local time when the sensor generates the sensing signal based on the sensor's local clock, and to obtain the sensor's correction time based on the difference between the receiving time and the sending time of the synchronization message received by the sensor from the central computing platform. The rolling force prediction device 30 includes a time correction module, used to correct the local time when the sensor generates the sensing signal based on the sensor's correction time, to obtain the global time when the sensor generates the sensing signal. The rolling force prediction device 30 includes an alignment and completion module, used to perform signal alignment and signal completion based on the global time when each sensor generates the sensing signal, to obtain sensor data at the same time. The sensor data at the same time are sequentially selected as the sensor data at the first time.

[0075] In some disclosed embodiments, signal completion is performed after signal alignment, and the sensor with the highest sampling frequency is used as the reference sensor, and all other sensors are used as sensors to be completed. The target time for signal completion between adjacent times of the sensors to be completed is determined by the global time of the actual sensing signal of the reference sensor between adjacent times, and signal completion is performed on the target time between adjacent times of the sensors to be completed based on cubic spline interpolation.

[0076] In some disclosed embodiments, the update frequency of the first weight and the second weight is synchronized with the clock of the vehicle chassis control domain; and / or, in the case that the third time is taken as the new first time after the current time reaches the first time, the step of performing feature extraction based on the sensor data of the first time to obtain the feature manifold of the first time is returned to update the first value of the first weight and the second value of the second weight at the new first time.

[0077] Please see Figure 4 , Figure 4 This is a schematic diagram of the framework of an embodiment of the driving control device of this application. The driving control device 40 includes: a current prediction module 41 and a torque control module 42. The current prediction module 41 is used to predict the rolling force at the current moment based on the sensor data at the current moment; wherein, the predicted rolling force at the current moment is obtained based on the above-described rolling force prediction device embodiment; the torque control module 42 is used to perform torque control on the vehicle motor based on the predicted rolling force at the current moment.

[0078] In the above scheme, the driving control device 40 predicts the rolling force based on the sensor data at the current moment. This predicted rolling force is obtained from the rolling force prediction device embodiment described above. Based on this predicted rolling force, torque control is then applied to the vehicle motor. On one hand, the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, while the Gaussian kernel captures local instantaneous jumps. This helps to maximize adaptability to complex operating conditions. Furthermore, since only the feature manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, the computational complexity of rolling force prediction is reduced. On the other hand, since the first value of the first weight and the second value of the second weight are adaptively determined based on the feature manifold extracted from the sensor data, and since the sliding variance of the feature manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can determine which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the weight may be adjusted based on the first weight. When the data feedback from the sensor fluctuates relatively little, the local variability of the feature manifold is low, resulting in a small sliding variance. Therefore, the first value of the first weight and the second value of the second weight can be adaptively increased and decreased. This allows the rolling force prediction model to focus more on the polynomial kernel to accurately grasp the long-term trend. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensor fluctuates at high frequencies, and the feature variance increases rapidly. Therefore, the adaptive weights will react immediately, that is, the first value of the first weight will be adaptively decreased and the second value of the second weight will be adaptively increased. This allows the rolling force prediction model to switch to focusing more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted according to the actual working conditions, which helps to improve the accuracy of rolling force prediction. On the other hand, when calculating the feature variance, the feature manifold at the first moment and the feature manifold at the second moment before the first moment are combined to calculate the sliding variance within the target sliding window. This can filter instantaneous sampling noise and help to ensure that the evolution of the weight factors between different working conditions is as smooth and continuous as possible, which also helps to improve the accuracy of rolling force prediction. Therefore, it can reduce the computational complexity of rolling force prediction and improve the accuracy of rolling force prediction, thereby helping to improve the response speed and accuracy of driving control.

[0079] Please see Figure 5 , Figure 5This is a schematic diagram of a framework of an embodiment of the electronic device of this application. The electronic device 50 includes at least a memory 51 and a processor 52 coupled to each other. The memory 51 stores at least program instructions, and the processor 52 is used to execute the program instructions to implement the steps in any of the above-described rolling force prediction method embodiments, or to implement the steps in any of the above-described driving control method embodiments. For details, please refer to the foregoing disclosed embodiments, which will not be repeated here. Exemplarily, the electronic device 50 may include, but is not limited to, industrial control computers, vehicle-mounted systems, etc., and the specific type of the electronic device 50 is not limited here.

[0080] Specifically, processor 52 controls itself and memory 51 to implement the steps in any of the above-described rolling force prediction method embodiments, or to implement the steps in any of the above-described driving control method embodiments. Processor 52 may also be referred to as a CPU (Central Processing Unit). Processor 52 may be an integrated circuit chip with signal processing capabilities. Processor 52 may also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component. A general-purpose processor may be a microprocessor or any conventional processor. Furthermore, processor 52 may be implemented using integrated circuit chips.

[0081] In the above scheme, the electronic device 50 performs feature extraction based on sensor data at the first moment to obtain the feature manifold at the first moment. Then, it analyzes the feature manifold at the first moment and the feature manifold at the second moment before the first moment to obtain the sliding variance of the feature manifold within the target sliding window. The target sliding window covers the first moment and the second moment before the first moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps. The first weight and the second weight are negatively correlated. Therefore, based on the sliding variance, the first value of the first weight at the first moment is determined, and based on the first value and the negative correlation, the second weight is determined. The rolling force prediction model focuses on the second value at the first moment, and the sliding variance is negatively correlated with the first weight. Based on this, the characteristic manifold at the first moment is predicted using the rolling force prediction model to obtain the predicted rolling force at the first moment. This is because the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with the first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, while the Gaussian kernel captures local instantaneous jumps, which helps to maximize adaptability to complex working conditions. Furthermore, since the prediction process only requires extracting the characteristic manifold and determining the first and second values ​​of the first and second weights, it also helps to reduce the computational complexity of rolling force prediction. On the other hand, based on the characteristic manifold extracted from sensor data... The model adaptively determines the first value of the first weight and the second value of the second weight. Since the sliding variance of the characteristic manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can determine which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the data feedback from the sensors fluctuates less, resulting in lower local variability of the characteristic manifold and thus a smaller sliding variance. Therefore, it can adaptively increase the first value of the first weight and decrease the second value of the second weight, allowing the rolling force prediction model to focus more on the polynomial kernel to accurately control long-term trends. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensors fluctuates at high frequencies. As the feature variance increases rapidly, the adaptive weights react immediately. The first value of the first weight decreases adaptively, and the second value of the second weight increases adaptively. This allows the rolling force prediction model to switch to focusing more on the Gaussian kernel to quickly capture local instantaneous changes. In this way, the weight factors can be quickly and sensitively adjusted to follow the actual working conditions, which helps to improve the accuracy of rolling force prediction. On the other hand, when calculating the feature variance, the sliding variance is calculated within the target sliding window by combining the feature manifold at the first moment and the feature manifold at the second moment before the first moment. This can filter out instantaneous sampling noise and help to ensure that the evolution of the weight factors between different working conditions is as smooth and continuous as possible, which also helps to improve the accuracy of rolling force prediction.Therefore, the computational complexity of rolling force prediction can be reduced, and the accuracy of rolling force prediction can be improved. In addition, torque control of the vehicle motor based on the predicted rolling force at the current moment helps to improve the response speed and accuracy of driving control.

[0082] Please see Figure 6 , Figure 6 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. The computer-readable storage medium 60 stores program instructions 61 that can be executed by a processor. The program instructions 61 are used to implement the steps in any of the above-described rolling force prediction method embodiments, or the steps in any of the above-described driving control method embodiments.

[0083] In the above scheme, the computer-readable storage medium 60 performs feature extraction based on sensor data at a first moment to obtain the feature manifold at the first moment. Then, it analyzes the feature manifold at the first moment and the feature manifold at a second moment prior to the first moment to obtain the sliding variance of the feature manifold within a target sliding window. The target sliding window covers the first moment and the second moment prior to the first moment. The rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with first and second weights, respectively. The polynomial kernel is responsible for the long-term trend of rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps. The first weight and the second weight are negatively correlated. Therefore, based on the sliding variance, the first value of the first weight at the first moment is determined, and based on the first value and the negative correlation, the rolling force prediction model is determined. The second weight is set to a second value at the first time step, and the sliding variance is negatively correlated with the first weight. Then, based on the rolling force prediction model, the characteristic manifold at the first time step is predicted to obtain the predicted rolling force at the first time step. On the one hand, since the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with the first and second weights respectively, and the polynomial kernel is responsible for the long-term trend of the rolling force while the Gaussian kernel is responsible for capturing local instantaneous jumps, it helps to maximize adaptability to complex working conditions. Furthermore, since only the characteristic manifold needs to be extracted during the prediction process to determine the first value of the first weight and the second value of the second weight, it also helps to reduce the computational complexity of rolling force prediction. On the other hand, since the rolling force prediction is based on data extracted from sensor data... The characteristic manifold adaptively determines the first value of the first weight and the second value of the second weight. Since the sliding variance of the characteristic manifold is negatively correlated with the first weight, and the first and second weights are negatively correlated, it can determine which kernel function the rolling force prediction model emphasizes at different times. For example, when driving smoothly on a structured road, the data feedback from the sensors fluctuates less, resulting in lower local variability of the characteristic manifold and thus a smaller sliding variance. Therefore, it can adaptively increase the first value of the first weight and decrease the second value of the second weight, allowing the rolling force prediction model to focus more on the multinomial kernel to accurately grasp the long-term trend. However, when encountering special road conditions such as speed bumps or potholes, the data feedback from the sensors fluctuates at high frequencies. At this point, the feature variance increases rapidly, so the adaptive weights will react immediately. The first value of the first weight will adaptively decrease and the second value of the second weight will adaptively increase. This allows the rolling force prediction model to switch to focusing more on the Gaussian kernel to quickly capture local instantaneous jumps. In this way, the weight factors can be quickly and sensitively adjusted according to the actual working conditions, which helps to improve the accuracy of rolling force prediction. On the other hand, when calculating the feature variance, the sliding variance is calculated within the target sliding window by combining the feature manifold at the first moment and the feature manifold at the second moment before the first moment. This can filter out instantaneous sampling noise and help to ensure that the evolution of the weight factors between different working conditions is as smooth and continuous as possible, which also helps to improve the accuracy of rolling force prediction.Therefore, the computational complexity of rolling force prediction can be reduced, and the accuracy of rolling force prediction can be improved. In addition, torque control of the vehicle motor based on the predicted rolling force at the current moment helps to improve the response speed and accuracy of driving control.

[0084] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0085] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus implementations described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0087] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0088] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the user through pop-up information or by asking the user to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A method for predicting rolling force, characterized in that, include: Feature extraction is performed based on the sensor data at the first moment to obtain the feature manifold at the first moment; Based on the characteristic manifold at the first time point and the characteristic manifold at the second time point before the first time point, the sliding variance of the characteristic manifold within the target sliding window is obtained; wherein, the target sliding window covers the first time point and the second time point before the first time point, and the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps, and the first weight and the second weight are negatively correlated; Based on the sliding variance, a first value of the first weight at the first time point is determined, and based on the first value and the negative correlation, a second value of the second weight at the first time point is determined; wherein the sliding variance is negatively correlated with the first weight. Based on the rolling force prediction model, the characteristic manifold at the first moment is predicted to obtain the predicted rolling force at the first moment.

2. The rolling force prediction method according to claim 1, characterized in that, Determining the first value of the first weight at the first time point based on the sliding variance includes: Obtain the variance difference between the sliding variance and the variance threshold; Obtain the reciprocal of the sum of the weight base with the natural constant as the base and the variance difference as the exponent, and 1, as the first value of the first weight at the first time.

3. The rolling force prediction method according to claim 1, characterized in that, The method further includes: Obtain the rolling force difference between the predicted rolling force and the actual rolling force at the first moment; In response to the rolling force difference exceeding the fault tolerance range, the feature manifold at the first moment is selected as the new support vector of the rolling force prediction model, and the correlation vector is obtained based on the feature distance between the new support vector and each of the old support vectors of the rolling force prediction model. The sensitivity matrix of the rolling force prediction model is updated based on the correlation vector to obtain a new sensitivity matrix of the rolling force prediction model; wherein, the sensitivity matrix reflects the sensitivity of the rolling force prediction model to data changes; Based on the new sensitivity matrix and the rolling force difference, the adjustment range of the weight parameters in the rolling force prediction model is obtained.

4. The rolling force prediction method according to claim 3, characterized in that, The correlation vector is obtained based on the feature distance between the new support vector and each of the old support vectors of the rolling force prediction model, including: Using the new support vector and any of the old support vectors as input parameters to the hybrid kernel function, the feature distance between the new support vector and the corresponding old support vector is obtained; The association vector is obtained based on the feature distances corresponding to each of the old support vectors.

5. The rolling force prediction method according to claim 3, characterized in that, The step of updating the sensitivity matrix of the rolling force prediction model based on the correlation vector to obtain a new sensitivity matrix for the rolling force prediction model includes: Based on the correlation vector and its transpose, a correction matrix is ​​obtained for updating the sensitivity matrix. The sensitivity matrix is ​​updated based on the product of the forgetting factor and the correction matrix to obtain the new sensitivity matrix.

6. The rolling force prediction method according to claim 1, characterized in that, The training steps for the rolling force prediction model include: The target hyperparameters for model training of the rolling force prediction model are obtained by searching the hyperparameter solution space of the rolling force prediction model. The rolling force prediction model is trained and optimized based on the sample training data under the control of the target hyperparameters. The absolute value of the difference between the predicted rolling force of the rolling force prediction model before and after performing the training optimization in this round is obtained as the stable value of the training optimization in this round. Based on the average of the stable values ​​obtained from several rounds of training optimization, it is determined whether to perform Cauchy mutation on the target hyperparameter to obtain a new target hyperparameter, and then re-execute the training optimization under the control of the new target hyperparameter.

7. The rolling force prediction method according to claim 6, characterized in that, The step of performing Cauchy mutation on the target hyperparameters to obtain new target hyperparameters includes: The scaling factor is obtained by multiplying the step size parameter by a random number that follows a Cauchy distribution. The target hyperparameters are scaled based on the scaling factor to obtain the new target hyperparameters.

8. The rolling force prediction method according to claim 6, characterized in that, The target hyperparameters include at least one of the following hyperparameters: penalty factor, insensitive loss coefficient, and weight ratio of the hybrid kernel function; And / or, the target hyperparameters are obtained by searching the hyperparameter solution space based on PSO-BAS; And / or, if the average value of the stable value optimized in several rounds of training is lower than a preset threshold, it is determined that the Cauchy mutation be performed on the target hyperparameter to obtain the new target hyperparameter; And / or, the sample training data includes sensor data at each sample time, the sensor data at each sample time is labeled with the actual rolling force, and the training loss of the rolling force prediction model is obtained by measuring the difference between the predicted rolling force and the actual rolling force.

9. The rolling force prediction method according to claim 1, characterized in that, The sensor data includes sensing signals from several sensors. The feature extraction based on the sensor data at the first moment to obtain the feature manifold at the first moment includes: Based on the sensed signal, the signal is decomposed to obtain several signal components and the center frequency of each signal component; Based on the correlation coefficient between the signal component and the sensed signal, it is determined whether to retain the signal component as a candidate component; Based on whether the center frequency of the candidate component is within the dynamic range, determine whether to retain the candidate component as the target component; Feature extraction is performed based on the data points at the first time point to obtain the feature manifold at the first time point; wherein, the data points at the first time point include the target component of the sensor.

10. The rolling force prediction method according to claim 9, characterized in that, The step of extracting features based on the data points at the first time point to obtain the feature manifold at the first time point includes: An objective function is constructed based on the difference between the weighted sum of the data points at each second time point and the data points at the first time point, using the weighting coefficients at each second time point respectively; wherein, the weighting coefficients are the unknowns of the objective function. The weighting coefficients obtained from solving the objective function at each second time point, and the feature vectors mapped to the low-dimensional space for each data point at each second time point, are used to obtain the feature vectors mapped to the low-dimensional space for the data points at the first time point; wherein, the feature vectors mapped to the low-dimensional space for the data points at the second time point are obtained during the rolling force prediction process at the second time point; Based on the sub-feature vectors belonging to each sensor in the feature vector at the first moment, the phase space is reconstructed to obtain the phase space trajectory of each sensor at the first moment. Based on the phase space trajectories of each of the sensors at the first moment, the characteristic manifold at the first moment is obtained.

11. The rolling force prediction method according to claim 1, characterized in that, The sensor data includes sensing signals from several sensors. Before performing feature extraction based on the sensor data at the first time moment to obtain the feature manifold at the first time moment, the method further includes: Based on the local clock of the sensor, the local time when the sensor generates the sensing signal is determined, and the correction time of the sensor is obtained based on the difference between the time when the sensor receives the synchronization message from the central computing platform and the time when the synchronization message is sent. The local time when the sensor generates the sensing signal is corrected based on the correction time of the sensor to obtain the global time when the sensor generates the sensing signal. Signal alignment and signal completion are performed based on the global time when each of the sensors generates the sensing signal to obtain sensor data at the same time; wherein, the sensor data at the same time is sequentially selected as the sensor data at the first time.

12. The rolling force prediction method according to claim 11, characterized in that, The signal completion is performed after the signal alignment, and the sensor with the highest sampling frequency is used as the reference sensor. All other sensors are used as sensors to be completed. The target time for the signal completion of the sensors to be completed between adjacent times is determined by the global time in which the sensing signal of the reference sensor actually exists between adjacent times. The signal completion is performed on the target time between adjacent times of the sensors to be completed based on cubic spline interpolation.

13. The rolling force prediction method according to any one of claims 1 to 12, characterized in that, The update frequency of the first weight and the second weight is synchronized with the clock of the vehicle chassis control domain; And / or, in the case of a third time after the current time reaches the first time, the third time is taken as the new first time, and the step of performing feature extraction based on the sensor data of the first time to obtain the feature manifold of the first time is returned to update the first value of the first weight at the new first time and the second value of the second weight at the new first time.

14. A driving control method, characterized in that, include: The predicted rolling force at the current moment is obtained by making a prediction based on the sensor data at the current moment; wherein the predicted rolling force at the current moment is obtained based on the rolling force prediction method according to any one of claims 1 to 13. Based on the predicted rolling force at the current moment, torque control is performed on the vehicle motor.

15. A rolling force prediction device, characterized in that, include: The feature extraction module is used to extract features based on the sensor data at the first time step to obtain the feature manifold at the first time step. The manifold analysis module is used to analyze the characteristic manifold based on the first time moment and the characteristic manifold at the second time moment before the first time moment to obtain the sliding variance of the characteristic manifold within the target sliding window; wherein, the target sliding window covers the first time moment and the second time moment before the first time moment, and the rolling force prediction model includes a hybrid kernel function that weights a polynomial kernel and a Gaussian kernel with a first weight and a second weight, respectively. The polynomial kernel is responsible for the long-term trend of the rolling force, and the Gaussian kernel is responsible for capturing local instantaneous jumps, and the first weight and the second weight are negatively correlated; The weight update module is used to determine a first value of the first weight at the first time based on the sliding variance, and to determine a second value of the second weight at the first time based on the first value and the negative correlation; wherein the sliding variance is negatively correlated with the first weight. The rolling prediction module is used to predict the characteristic manifold at the first moment based on the rolling force prediction model to obtain the predicted rolling force at the first moment.

16. A driving control device, characterized in that, include: The current prediction module is used to make a prediction based on the sensor data at the current moment to obtain the predicted rolling force at the current moment; wherein the predicted rolling force at the current moment is obtained based on the rolling force prediction device according to claim 15; The torque control module is used to control the torque of the vehicle motor based on the predicted rolling force at the current moment.

17. An electronic device, characterized in that, It includes at least a memory and a processor coupled to each other, the memory storing at least program instructions, and the processor executing the program instructions to implement the rolling force prediction method according to any one of claims 1 to 13, or the driving control method according to claim 14.

18. A computer-readable storage medium, characterized in that, The device stores program instructions that can be executed by a processor, the program instructions being used to implement the rolling force prediction method according to any one of claims 1 to 13, or the driving control method according to claim 14.

Images