Automatic parameter adjusting method based on KPI result and domain controller

Abstract

The invention relates to an automatic parameter adjusting method based on a KPI result. The method includes: performing KPI item calculation related to driving control based on an input vehicle measurement signal and outputting a KPI result; adjusting and optimizing the driving control parameters according to the KPI result and outputting the optimal driving control parameters; and calculating a control instruction according to the optimal driving control parameter and outputting the control instruction. According to the invention, the parameter adjustment efficiency can be effectively improved, and the influence of the driving experience and driving scene of the driver can be avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the automatic driving technology, and in particular to an automatic parameter tuning method based on KPI results and a domain controller. BACKGROUND

[0002] In the current development of ADAS (i.e. Advanced Driver Assistance System) or AD (i.e. Autonomous Driving System), manual tuning of driving control parameters (also referred to as "parameter tuning" for short) is usually required to enable the autonomous driving system or the driving assistance system to achieve good control.

[0003] However, the workload required in the parameter tuning process will vary depending on the experience of the application engineers engaged in the parameter tuning, and since the quality of the control performance often depends on subjective evaluation rather than objective criteria, this will lead to multiple rounds of manual parameter tuning to meet the requirements of the demand side. SUMMARY

[0004] Based on the above problems in the prior art, the present application aims to provide an automatic parameter tuning method based on KPI results, which can effectively improve the parameter tuning efficiency.

[0005] The automatic parameter tuning method based on KPI results according to an aspect of the present application comprises:

[0006] calculating KPI items related to driving control based on the input vehicle measurement signals and outputting KPI results;

[0007] tuning driving control parameters according to the KPI results and outputting optimal driving control parameters; and

[0008] calculating control instructions according to the optimal driving control parameters and outputting the control instructions. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 is a structural block diagram of an automatic parameter tuning system based on KPI results according to an embodiment of the present application.

[0010] Figure 2 is a schematic diagram showing an example of implementing an LQR algorithm in a control module.

[0011] Figure 3 is a schematic diagram showing the centered response performance.

[0012] Figure 4 is a schematic diagram showing an example of an automatic parameter tuning module.

[0013] Figure 5 is a schematic diagram of the automatic parameter tuning process using the centered performance (Pdy) as an example.

[0014] Figure 6 is a schematic diagram of the machine learning mechanism in the embodiment.

[0015] Figure 7 One example of the machine learning model is illustrated.

[0016] Figure 8 is a flowchart illustrating the automatic parameter tuning method based on KPI results of the present application. DETAILED DESCRIPTION

[0017] The following describes some of the embodiments of the present application, which are intended to provide a basic understanding of the present application and are not intended to identify key or decisive elements or limit the scope of protection.

[0018] Figure 1 is a structural block diagram of the automatic parameter tuning system based on KPI results of an embodiment of the present application.

[0019] As shown in Figure 1 , the automatic parameter tuning system based on KPI results of an embodiment of the present application comprises:

[0020] a control module 100 for calculating and outputting control instructions based on driving control parameters;

[0021] a KPI evaluation module 200 for calculating and outputting KPI results based on vehicle measurement signals inputted in relation to driving control; and

[0022] an automatic parameter tuning module 300 for tuning driving control parameters according to the KPI results and outputting optimal driving control parameters to the control module 100.

[0023] Among them, the control module 100 may be, for example, one of the software control modules in the ADAS (Advanced Driver Assistance System) or AD (Automatic Driving System), which is used to realize vehicle driving control in different scenarios.

[0024] As shown in Figure 1 , the control module 100 provides the control instruction C1 to the vehicle, the vehicle controls the driving of the vehicle according to the control instruction C1, the KPI evaluation module 200 is used to obtain the measurement signal M1 of the vehicle and output the KPI result P after processing to the automatic parameter tuning module 300, the automatic parameter tuning module 300 tunes the parameters according to the KPI result P and outputs the optimal parameters K to the control module 100, and the control module 100 calculates the control instruction C1 based on the optimal parameters K and provides it to the vehicle.

[0025] In the control module 100, there are generally three control algorithms in vogue, PID algorithm, LQR algorithm (or also a combination of LQR algorithm and PID algorithm) and MPC algorithm, for different control algorithms, there are different weight parameters that need to be calibrated, at the same time, there are also some filtering parameters or estimation parameters that need to be calibrated, these parameters are also collectively referred to as "driving control parameters" in the present application.

[0026] In the following description, the LQR algorithm will be taken as an example for description.

[0027] In the present embodiment, first, it is explained how to use the automatic parameter tuning method based on KPI results of the present application in the lateral control of the ADAS system, wherein the LQR algorithm is used to implement the lateral control in the control module 100.

[0028] Figure 2 is a schematic diagram representing an example of implementing the LQR algorithm in the control module.

[0029] The control module 100 includes a feedforward control module 110 and a feedback control module 120. The control module 100 receives the target trajectory and the vehicle state provided by the perception module and the sensor, which are used to estimate the state error of the LQR control model (the state error includes: lateral deviation, lateral deviation rate, heading angle and heading angle rate), and provide the target trajectory and the vehicle state to the feedforward control module 110 and the feedback control module 120 respectively for feedforward control and feedback control and output the feedforward control result and the feedback control result, the sum of the feedforward control result and the feedback control result will constitute the final EPS (Electric Power Steering) steering angle.

[0030] Specifically, as for the feedforward control module 110, its role is to adjust the direction of the vehicle in advance to reduce the tracking error, as shown in Figure 2 The feedforward control module 110 mainly performs the following steps:

[0031] a3: obtaining the target curvature radius of the trajectory according to the target trajectory a1 and the vehicle state a2;

[0032] a4: performing steady-state steering angle calculation according to the target curvature radius (the formula in a4 represents the formula for calculating the steady-state steering angle, which belongs to the prior art);

[0033] a5: performing steady-state direction angle error compensation calculation according to the target curvature radius and one item k3 from the feedback control law matrix K of the feedback control module 120 (the formula in a5 represents the formula for calculating the steady-state direction angle error compensation, which belongs to the prior art);

[0034] a6: The sum of the steady-state steering angle and steady-state direction angle error is taken as the feedforward angle (the formula in a6 represents the formula for calculating the feedforward angle, which belongs to the prior art).

[0035] The inputs to the control module 100, namely trajectory information and vehicle status, employ a PT1 filter and look-ahead point setting logic. The key parameters are as follows:

[0036] KtiLookAhead: Used to calculate forward-looking distances;

[0037] KtiPt1_curvature_ffw: A PT1 filter (i.e., a first-order low-pass filter) used for feedforward control of the trajectory curvature signal.

[0038] KtiPt1_dy_and_ag: A PT1 filter used for state error in feedback control.

[0039] For the feedforward control module 110, the radius of curvature of the trajectory is crucial for determining the final feedforward angle. Therefore, the main parameter is the PT1 filter time constant KtiPt1_curvature_ffw. This is because, in the feedforward control module 110, the radius of curvature of the trajectory is one of the key factors in calculating the final feedforward angle (such as the steering angle). The radius of curvature information directly affects the magnitude and direction of the steering action that the vehicle needs to take to ensure that the vehicle can travel along the predetermined trajectory. The PT1 filter time constant KtiPt1_curvature_ffw is used to smooth the radius of curvature signal to reduce control instability that may be caused by measurement noise or signal abrupt changes. By adjusting this time constant, the response speed and filtering effect of the filter can be controlled.

[0040] Furthermore, regarding the feedback control module 120, such as Figure 2 As shown, the feedback control module 120 mainly performs the following steps:

[0041] a7: Based on the target trajectory a1 and vehicle state a2, obtain the vehicle dynamics model and steering geometry model, and based on the vehicle dynamics model and steering geometry model, obtain the state space equation;

[0042] a8: Perform LQR optimization control based on the state-space equation and obtain the feedback control law matrix K. In this case, a term k3 in the feedback control law matrix K is provided to the steady-state direction angle compensation calculation in a5. Here, k3 represents a term in the feedback control law matrix K = [k1 k2k3k4] obtained by solving the feedback of LQR.

[0043] a9: Calculate the optimal front wheel steering angle u based on the feedback control law matrix K.

[0044] Finally, in the control module 100, such as Figure 2 As shown in a10, the feedforward angle output by the feedforward control module 110 and the optimal front wheel steering angle u output by the feedback control module 120 are added together to output the final steering angle.

[0045] The feedback control in the feedback control module 120 relies on the actual output of the system to adjust the control input to reduce errors. In autonomous driving, feedback control compares the difference between the actual state of the vehicle (such as position and orientation) and the target state, and then adjusts the control input such as the front wheel steering angle based on this difference to reduce errors and improve the vehicle's tracking performance.

[0046] For example, adjusting the Q and R parameters of the LQR algorithm to obtain an optimized solution ( Figure 2 In case of a8), the feedback request value will be given based on the optimized solution value of LQR. The parameter Q includes: Q = [Q0, Q1, Q2, Q3]: Q0 is the penalty weight for state dy (lateral deviation error), Q1 is the penalty weight for state dyRate (lateral deviation rate of change error), Q2 is the penalty weight for state headingAg (heading angle error), and Q3 is the penalty weight for state headingAgRate (heading angle rate of change error); R is the penalty weight of the output.

[0047] On the other hand, PID algorithms employ integral controllers. Typically, the integral controller (i.e., the I-term control in a PID algorithm) is primarily used to eliminate the system's steady-state error. For example, when a vehicle is traveling on a straight road with a cross slope, it will travel with a constant lateral deviation from the lane centerline. In this case, the steady-state error can be adjusted by modifying the Ki factor of the integral controller (i.e., the integral gain coefficient of the lateral deviation error).

[0048] Moreover, in some complex control scenarios, in order to balance the steady-state performance and dynamic performance of the system, the integral controller may be used in combination with the LQR controller. For example, the integral term of the integral controller can be added to the LQR controller to further eliminate steady-state error.

[0049] in addition, Figure 2 In this equation, b1 represents the vehicle dynamics model and steering geometry model, b2 represents the state-space equation, b3 represents the objective function formula related to LQR optimal control, and b4 represents the formula for calculating the optimal front wheel steering angle u. Figure 2 The specific contents and related formulas of the feedforward control module 110 and the feedback control module 120 are all existing technologies and will not be described in detail here.

[0050] Next, the KPI (Key Performance Indicator) evaluation module 200 will be explained.

[0051] KPI evaluation module 200 aims to calculate KPI results for control performance. It assumes that the signals measured from sensors, sensing modules, or fusion modules are absolutely accurate and ground-true. Therefore, the KPI results of this evaluation module 200 only refer to control module 100 and do not involve the entire system level. KPI evaluation module 200 focuses on evaluating the performance of control module 100, particularly its efficiency and accuracy in handling longitudinal and lateral control tasks. Under the assumption of absolute accuracy of sensor data, this evaluation module 200 can more directly focus on the performance of the control algorithm, providing important basis for system optimization and debugging.

[0052] The KPI evaluation module 200 covers all KPI items used to indicate control performance in all driving scenarios. For example, KPI items related to longitudinal control include acceleration / deceleration shock, cut-in / cut-out reaction time, and approach distance. KPI items related to lateral control include centering control (lateral deviation control), centering response, and curve traversal capability.

[0053] The following are some KPI items that can be defined and calculated in KPI evaluation module 200:

[0054] (1) Centering performance (P) dy )

[0055] Definition: The deviation in distance between the vehicle center and the center of the lane. Centering performance will record and calculate the average, peak, standard deviation, and proportions of lateral deviation within ±10cm, ±15cm, and ±20cm.

[0056] As an example, centering performance (P) dy The calculation process may include the following steps:

[0057] Lane line recognition: This typically involves acquiring road images or point cloud data through sensors such as cameras or LiDAR, and then using image processing or machine learning algorithms to extract lane line information.

[0058] Vehicle position determination: Real-time determination of the vehicle's position, including its lateral position (i.e., the distance between the vehicle's center and the lane centerline) and longitudinal position, is typically achieved through GPS, inertial navigation systems (INS), or methods fusing data from multiple sensors; and

[0059] KPI Calculation: Based on the information obtained from the above two steps, calculate the average, peak, and standard deviation of the lateral deviation, as well as the proportion of dy deviations within the ranges of ±10cm, ±15cm, and ±20cm.

[0060] Among them, the average value of lateral deviation is calculated by recording the lateral distance (deviation) between the vehicle center and the lane centerline in real time or periodically and calculating the average value of these deviations. This average value reflects the vehicle's ability to remain parallel to the lane centerline during driving and is an important indicator for measuring vehicle driving stability and safety.

[0061] Regarding peak values, the maximum and minimum values ​​of lateral deviation are recorded and calculated, i.e., peak values, which can represent a large deviation of the vehicle in a short period of time.

[0062] Regarding standard deviation, the standard deviation of lateral deviation is calculated to assess the fluctuation of vehicle offset during driving. The larger the standard deviation, the more unstable the vehicle driving.

[0063] Regarding the proportion of dy deviation within a specific range, the proportion of dy deviation within ±10cm, ±15cm and ±20cm is calculated respectively. These proportions can intuitively reflect the vehicle's performance under different degrees of offset.

[0064] Furthermore, this KPI project can also include different sub-indicators for different scenarios, such as centering performance P on straight roads (curvature radius > 5000m). dy The centering performance Pdy on a right turn curve with R=500m and P on a left turn curve with R=200m dy wait.

[0065] (2) Centered response performance (P) rec )

[0066] Definition: When the lane centering function is activated as the vehicle approaches the right / left lane of the lane, the vehicle returns to the center of the lane and reaches a steady state.

[0067] The centering response performance mainly has three sub-indicators: maximum overshoot distance P rec_maxdist Centering response time (i.e., the time required for the vehicle to steer back to near the center of the lane and reach a steady state) P rec_time and the number of oscillations P during this period rec_oc Simultaneously, this KPI will be recorded and calculated for different scenarios, such as centering performance P. dy .

[0068] Figure 3 This is a schematic diagram illustrating the centering response performance. Figure 3 In the diagram, A represents the maximum overshoot value, t1 represents the homing time, t2 represents the convergence time, and T represents the homing response time.

[0069] It can be seen that the centering response performance (P) recKPIs are key indicators for measuring the efficiency and stability of a vehicle returning to the center of the lane and reaching a steady state after activating the lane centering function when approaching the lane lines. The following lists the calculation methods for the main KPI items used to calculate centering response performance:

[0070] ① Maximum overshoot distance (P) rec_maxdist ):

[0071] Definition: The maximum lateral distance between a vehicle and the lane centerline as the vehicle moves toward the center of the lane after the lane centering function is activated.

[0072] KPI calculation: By recording the lateral position data of the vehicle after activating the lane centering function, find the maximum deviation value between the vehicle and the lane center line.

[0073] ② Rebound response time (P) rec_time ):

[0074] Definition: The time required from when the vehicle activates the lane centering function until the vehicle returns to the vicinity of the lane center and reaches a steady state.

[0075] Calculation method: Record the activation time and the time when the vehicle reaches a steady state (i.e., the lateral deviation is within an acceptable range and no longer fluctuates significantly). The time difference between the two is the centering response time.

[0076] ③ Number of oscillations (P) rec_oc ):

[0077] Definition: The number of times the vehicle's lateral position fluctuates (oscillates) around the lane centerline during the centering response process.

[0078] Calculation method: By analyzing lateral position data, identify and count the number of periodic fluctuations that occur before the vehicle returns to the center of the lane and reaches a steady state.

[0079] As an example, the calculation process for the above key performance indicators (KPIs) includes the following steps:

[0080] Data collection: Use onboard sensors (such as cameras, radar, lidar, etc.) to acquire real-time lateral position data of the vehicle and record the time when the vehicle activates the lane centering function;

[0081] Data processing: The collected lateral position data is filtered and smoothed to eliminate noise and outliers, and key event points are extracted, such as the maximum overshoot point and the time point when the steady state is reached.

[0082] KPI Calculation: Based on the definitions and calculation methods, calculate the maximum overshoot distance, homing response time, and number of oscillations.

[0083] (3) Curve performance (P) curpass )

[0084] Definition: Vehicle performance through curves, including maximum response speed on different curves, entry / exit performance on curves, and centering performance while driving on curves. Specifically:

[0085] ① Maximum speed response: Through simulation or experimental testing, evaluate the maximum safe driving speed of the vehicle on different curves (such as different radii, different slopes, etc.).

[0086] ②Entry / exit curve performance: Evaluate the vehicle's stability, steering response speed, and accuracy when entering and exiting curves. This can be assessed by measuring indicators such as trajectory deviation and speed change when entering and exiting curves.

[0087] ③ Centering performance when driving on curves: Evaluate the vehicle's ability to maintain the center line of the lane when driving on curves.

[0088] (4) Vehicle stability (P) vehstable )

[0089] Definition: This KPI can be represented by unexpected oscillations and shocks in lateral acceleration. It can be calculated based on the yaw rate and lateral acceleration values ​​from an IMU device (IMU refers to an inertial measurement unit, i.e., from other ECUs). This KPI will include the duration and maximum value of the oscillation. The duration of the oscillation refers to the length of time the vehicle oscillates, usually expressed in seconds (s). The maximum value of the oscillation refers to the maximum lateral acceleration value reached by the vehicle during the oscillation, usually expressed in meters per second squared (m / s²).

[0090] As an example, the method for calculating vehicle stability includes the following steps:

[0091] Data acquisition: Yaw rate and lateral acceleration data are collected in real time using IMU devices installed on the vehicle;

[0092] KPI project calculation: The collected data is filtered to remove noise and interference signals, improve the accuracy and reliability of the data, and analyze the oscillation phenomenon in the data, including the duration, frequency, amplitude and other characteristics of the oscillation.

[0093] (5) Steering stability (P) strewhl )

[0094] Definition: This KPI can be represented by unexpected steering oscillations and can be calculated based on the steering speed of the ECU of EPS (Electric Power Steering) / SAS (Stability Control System).

[0095] As described above, the KPI evaluation module 200 calculates one or more KPI items as described in (1) to (5) above based on driving environment data (e.g., lines, objects, etc.) and vehicle state data (e.g., speed, steering angle, lateral acceleration, etc.) obtained from the vehicle's ADAS system, obtains KPI results, and outputs them to the automatic parameter tuning module 300.

[0096] As described above, in this invention, the KPI evaluation module 200 specifically defines objective goals (i.e., multiple KPI items) and calculates KPI results for these KPI items. Therefore, by providing such an objective evaluation expressed in KPI results, it is less susceptible to the influence of the driver's driving experience and driving scenario.

[0097] Next, the automatic parameter tuning module 300 will be explained.

[0098] The automatic parameter tuning module 300 optimizes the parameters based on the KPI results output by the KPI evaluation module 200 and outputs the optimal parameters to the control module 100. The control module 100 calculates control commands based on the optimal parameters and provides them to the vehicle for driving control.

[0099] In this embodiment, as an example, the automatic parameter tuning module 300 will run two main mechanisms: the first is an expert experience-based mechanism, which is also the only mechanism running in the initial stage. The concept of the expert experience-based mechanism is to realize the manual parameter tuning steps of experienced application engineers. For the expert experience-based mechanism in the automatic parameter tuning module 300, the main task is to turn the manual parameter tuning operation into an automatic operation mechanism; the second is a machine learning-based mechanism, which can be implemented using decision tree algorithms or reinforcement learning.

[0100] Figure 4 This is a schematic diagram illustrating an example of an automatic parameter tuning module.

[0101] exist Figure 4 In the diagram, region A1 represents the expert-based mechanism, and region A2 represents the machine learning-based mechanism. The expert-based mechanism, i.e., the parameter tuning data and corresponding KPI results generated in region A1 each time, will be fed back to the machine learning-based mechanism, i.e., region A2, for training. Figure 4 (As shown by arrow A12 in the diagram), on the other hand, the recommended parameters K generated after training in region A2 will be provided to step S5 in region A1 (…). Figure 4 (As indicated by arrow A21 in the diagram).

[0102] Figure 4 The A1 area process includes:

[0103] Step S1: Receive the current KPI result P;

[0104] Step S2: Determine whether the current KPI result P is better than the target KPI value C_tar. If the result is yes (Y), continue to step S3; otherwise, jump to step S4.

[0105] Step S3: Determine whether the current KPI result P is better than the previous KPI value P_last. If the result is yes (Y), continue to step S5; otherwise, skip to step S6.

[0106] Step S4: Keep parameter K constant;

[0107] Step S5: Update the parameter K in a prescribed manner to make it the recommended parameter K generated after training in the A2 region. Here, the prescribed update methods can be listed as: fixed step size update, dynamic step size update and direct update. Furthermore, it is preferred to use a small step size for update. This is to avoid the sudden performance drop that may be caused by a large change in parameter K, that is, to gradually adjust parameter K to try to improve performance.

[0108] Step S6: Reset parameter K to its previous value;

[0109] Step S7: Output parameter K.

[0110] The following will use the centering performance Pdy parameter tuning workflow as an example to illustrate the process.

[0111] Figure 5 With the centralization performance (P) dy A schematic diagram of the automatic parameter tuning process, taking as an example.

[0112] Figure 5 The left side includes the following steps:

[0113] Step S11: Receive the current KPI result P dy ;

[0114] Step S12: Determine the current P dy Is it better than the target KPI value C? dy_ta If the judgment result is (Y), continue to step S13; otherwise, jump to step S14.

[0115] Step S13: Determine the current P dy Is it better than the previous KPI value P? dy_last If the judgment result is yes (Y), continue to step S15; otherwise, skip to step S16.

[0116] Step S14: Keep parameter K constant;

[0117] Step S15: Update parameter K in the prescribed manner to become the recommended parameter K;

[0118] Step S16: Reset parameter K to its previous value;

[0119] Step S17: Output parameter K.

[0120] As an example, the recommended parameter K in step S15 is derived from... Figure 5 The steps shown on the right are provided as follows:

[0121] Step S20: Begin adjusting parameter K;

[0122] Step S21: Determine whether the road is straight (R>5000m). If the result is yes (Y), continue to step S22. If the result is no (N), skip to step S25.

[0123] Step S22: Adjust only the parameters of the feedback control;

[0124] Step S23: Determine whether the steady-state error needs to be adjusted. If the result is yes (Y), continue to step S24. If the result is no (N), jump to step S27.

[0125] Step S24: Adjust the Ki factor of the integral controller;

[0126] Step S25: Determine whether the feedforward output is sufficient (e.g., >85%). If the result is yes (Y), continue to step S26. If the result is no (N), jump to step S28.

[0127] Step S26: Determine whether the path tracking performance (Pstrewhl) is better than the target tracking performance (Cstrewhl_tar). If the result is yes (Y), continue to step S27. If the result is no (N), jump to step S29.

[0128] Step S27: Adjust the R and Q parameters of the linear quadratic regulator;

[0129] Step S28: Adjust the parameters of the PT1 filter for trajectory curvature;

[0130] Step S29: Adjust the parameters of the PT1 filter for state error.

[0131] It should be noted that the parameter K in the above description does not refer to a single parameter, but can actually include... Figure 2 The various control parameters involved in the control module 100, as an example, are, for instance, based on... Figure 5 As shown in the flowchart on the right, the parameter K can include: the Ki factor of the integral controller, the R and Q parameters of the linear quadratic regulator, the parameters of the PT1 filter for trajectory curvature, the parameters of the PT1 filter for state error, and so on.

[0132] Figure 6 This is a schematic diagram of the machine learning mechanism in this embodiment. Figure 6 The document describes the process of processing a dataset using two different machine learning models (decision trees and reinforcement learning) to obtain output results. Figure 6 The left side represents the dataset processing procedure using a decision tree model, and the right side represents the dataset processing procedure using a reinforcement learning model.

[0133] Figure 6 The dataset processing on the left side using a decision tree model includes:

[0134] Generate dataset 61 based on data R (data R includes: parameter tuning data and corresponding KPI results);

[0135] Input dataset 61 into decision tree model 62; and

[0136] The decision tree model 62 outputs parameter K after processing.

[0137] The dataset processing steps using the reinforcement learning model on the right include:

[0138] Generate dataset 71 based on data R (data R includes: parameter tuning data and corresponding KPI results);

[0139] Data set 71 is fed into reinforcement learning model 72;

[0140] The output parameter K is obtained after processing by reinforcement learning model 72.

[0141] Figure 7 This illustrates an example of a machine learning model.

[0142] Here's an example, assuming a simple and generally defined KPI: the lateral error P of a vehicle traveling at 100 kph on a 1000m radius curve. dy The corresponding KPI target is C. dy Among them, there are several parameters K dy =[k1,k2,k3…] is used to adjust this KPI item P dy Thus, as Figure 7 As shown, the input to machine model 80 is P. dy_v100_R1000 With C dy_v100_R1000 The difference and the previous parameter value K dy_last The output of machine model 80 is the updated K. dy (To be provided to ADAS or AD systems).

[0143] The above describes the automatic parameter tuning system based on KPI results of the present invention. Next, the automatic parameter tuning method based on KPI results of the present invention will be described.

[0144] Figure 8 This is a flowchart illustrating the automatic parameter tuning method based on KPI results of the present invention.

[0145] like Figure 8 As shown, the automatic parameter tuning method based on KPI results of the present invention corresponds to the automatic parameter tuning system based on KPI results of the present invention described above. This method mainly includes the following steps:

[0146] KPI calculation step S100: Calculate the KPI items related to driving control based on the input vehicle measurement signals and output the KPI results;

[0147] Automatic parameter tuning step S200: Based on the KPI results, optimize the driving control parameters and output the optimal driving control parameters; and

[0148] Control step S300: Control the driving of the vehicle according to the optimal driving control parameters.

[0149] In the automatic parameter tuning step S200, the driving control parameters are tuned based on the KPI results, the pre-set target KPI, and the KPI results of the previous driving control parameter tuning.

[0150] The automatic parameter tuning step S200 includes the following sub-steps:

[0151] Determine whether the KPI result is better than the pre-set target KPI;

[0152] If the KPI result is determined to be better than the preset target KPI, then it is further determined whether the KPI result is better than the KPI result of the previous driving control parameter optimization; otherwise, the current driving control parameters are kept unchanged.

[0153] If the KPI result is better than the KPI result of the previous driving control parameter optimization, the driving control parameters are updated with a specified step size; otherwise, the current driving control parameters are restored to the driving control parameters of the previous driving control parameter optimization.

[0154] The step of updating the driving control parameters with a specified step size includes:

[0155] The KPI results are input into a trained machine learning model, wherein the model is trained using multiple sets of training data. Each set of training data includes: the optimal driving control parameters obtained from each driving control parameter tuning and the corresponding KPI results for those optimal driving control parameters; and

[0156] The output information of the model is obtained as the driving control parameters updated with a specified step size, wherein the output information includes the optimal driving control parameters corresponding to the KPI results.

[0157] The automatic parameter tuning system and method based on KPI results of the present invention can be integrated into the autonomous driving or assisted driving system of the vehicle terminal, for example, integrated into the ADAS domain controller, or run as an independent software tool on a computer, connected to the ADAS control system ECU to participate in the vehicle performance tuning work, or integrated into independent controller hardware, or integrated into the cloud, and used to achieve vehicle performance tuning by connecting to the cloud.

[0158] The automatic parameter tuning system and method based on KPI results of this invention convert evaluation indicators into objective KPI values, thus avoiding the influence of driver experience and driving scenarios. Furthermore, by using mechanisms based on expert experience and machine learning for automatic parameter tuning, development efficiency is significantly improved compared to manual tuning, resulting in a substantial increase in system development efficiency and robustness of control performance. Moreover, through continuous learning and optimization based on machine learning, the vehicle control module can achieve continuous self-optimization.

[0159] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Those skilled in the art can conceive of other feasible variations or substitutions based on the technical scope disclosed in this application, and such variations or substitutions are all covered within the scope of protection of this application. Where there is no conflict, the embodiments and features described in the embodiments of this application can also be combined with each other. The scope of protection of this application is determined by the claims.

Claims

1. An automatic parameter tuning method based on KPI results, characterized in that, comprising the following steps: calculating KPI items related to driving control based on the input vehicle measurement signals and outputting KPI results; optimizing driving control parameters according to the KPI results and outputting optimal driving control parameters; and calculating control instructions according to the optimal driving control parameters and outputting the control instructions.

2. The automatic parameter tuning method based on KPI results according to claim 1, wherein the optimizing driving control parameters according to the KPI results and outputting optimal driving control parameters comprises: optimizing current driving control parameters based on the KPI results, pre-set target KPIs and KPI results of previous driving control parameter optimization.

3. The automatic parameter tuning method based on KPI results according to claim 1, wherein the optimizing driving control parameters according to the KPI results and outputting optimal driving control parameters comprises: judging whether the KPI results are better than pre-set target KPIs; in the case that the KPI results are better than pre-set target KPIs, further judging whether the KPI results are better than KPI results of previous driving control parameter optimization, otherwise keeping the current driving control parameters unchanged; in the case that the KPI results are better than KPI results of previous driving control parameter optimization, updating driving control parameters in a specified manner and taking them as the optimal driving control parameters, otherwise restoring the current driving control parameters to driving control parameters of previous driving control parameter optimization and taking them as the optimal driving control parameters.

4. The automatic parameter adjustment method based on KPI results according to claim 3, wherein, the updating driving control parameters in a specified manner comprises: inputting the KPI results into a trained machine learning model, wherein the model is trained using multiple sets of training data, each of which includes optimal driving control parameters obtained by each driving control parameter optimization and KPI results corresponding to the optimal driving control parameters; obtaining output information of the machine learning model, wherein the output information includes optimal driving control parameters corresponding to the KPI results; and updating driving control parameters in the specified manner until they become the optimal driving control parameters corresponding to the KPI results.

5. The automatic parameter tuning method based on KPI results according to claim 4, wherein the specified manner updating includes any one of the following manners: fixed step updating, dynamic step updating and direct updating.

6. The automatic parameter adjustment method based on KPI results according to claim 4, wherein, the trained machine learning model is obtained by: obtaining a training data set, which includes optimal driving control parameters obtained by each driving control parameter optimization and KPI results corresponding to the optimal driving control parameters; selecting a machine learning model, which is selected from a decision tree model or a reinforcement learning model; and training the machine learning model based on the training data set to obtain the trained machine learning model.

7. The automatic parameter tuning method based on KPI results according to claim 1, wherein The KPI result based on the input vehicle measurement signal includes: a KPI item related to lateral control is calculated according to the vehicle measurement signal, and a corresponding KPI value is obtained as the KPI result; a KPI item related to longitudinal control is calculated according to the vehicle measurement signal, and a corresponding KPI value is obtained as the KPI result.

8. The automatic parameter tuning method based on KPI result according to claim 1, wherein: the KPI item includes any one or more of the following: centering control performance, centering response performance, curve performance, vehicle stability, steering stability, acceleration impact, deceleration impact, cut-in reaction time, cut-out reaction time, and proximity distance.

9. The automatic parameter adjustment method based on KPI results according to claim 1, wherein, the calculation of the control instruction according to the optimal driving control parameter includes any one of the following: an LQR algorithm is used to calculate the control instruction; a PID algorithm is used to calculate the control instruction; and an LQR algorithm and a PID algorithm are combined to calculate the control instruction.

10. A domain controller, characterized by a storage module, a processor, and a computer program stored on the storage module and executable on the processor, wherein the processor implements the automatic parameter tuning method based on KPI result according to any one of claims 1-9 when executing the computer program.

11. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the automatic parameter tuning method based on KPI result according to any one of claims 1-9.

Images