Defect repairing method and device for automatic driving function model, equipment and medium

Abstract

The invention discloses a defect repairing method and device for an automatic driving function model, equipment and a storage medium, and the method comprises the steps: determining a constraint condition corresponding to the automatic driving function model, and first output data of the automatic driving function model; determining a defect detection result of the automatic driving function model based on the first output data and constraint conditions; in response to the defect detection result that the automatic driving function model has defects, determining first input data corresponding to the automatic driving function model based on the first output data; and repairing the defect of the automatic driving function model based on the first input data and the constraint condition. According to the technical scheme, defect detection of the model is achieved through the output data of the model and the constraint conditions, defect repair of the model is achieved through the input data corresponding to the output data and the constraint conditions, and therefore defect detection and repair of the model can be completed without depending on a defect specific form, and the defect repair efficiency is improved. Therefore, the safety and reliability of the model are enhanced.

Description

Technical Field

[0001] The present disclosure relates to the field of intelligent driving, and in particular to a defect repair method, device, equipment, and medium for an autonomous driving function model. Background Art

[0002] In recent years, with the development of intelligent driving technology, assisted driving functions or automatic driving functions have become more and more popular among consumers. As a result, the safety of intelligent driving technology has also attracted increasing attention.

[0003] At present, intelligent driving systems (such as assisted driving systems and autonomous driving systems) are embedded with a large number of software, communication channels, sensors and other information processing units, making them more vulnerable to cyber and physical attacks. For example, sensor spoofing attacks have become a prominent problem in intelligent driving systems, where hackers can manipulate sensor measurements arbitrarily, thereby undermining system security or driving the system to produce unsafe behaviors. Summary of the invention

[0004] At present, in order to ensure the safety of intelligent driving systems, the commonly used solution is a defect detection method based on hardware constraints, but this method can only detect defects but cannot repair them. In order to solve the above technical problems, the present disclosure provides a defect repair method and device, equipment, and medium for an autonomous driving function model, which can repair detected defects in a timely manner.

[0005] According to a first aspect of the present disclosure, a method for repairing defects in an autonomous driving function model is provided, comprising: determining constraints corresponding to the autonomous driving function model, and first output data of the autonomous driving function model; determining a defect detection result of the autonomous driving function model based on the first output data and the constraints; in response to the defect detection result indicating that the autonomous driving function model has defects, determining first input data corresponding to the autonomous driving function model based on the first output data; and repairing the defects of the autonomous driving function model based on the first input data and the constraints.

[0006] According to a second aspect of the present disclosure, a defect repair device for an autonomous driving function model is provided, comprising: an output data determination module, used to determine the constraint conditions corresponding to the autonomous driving function model, and first output data of the autonomous driving function model; a model defect detection module, used to determine the defect detection result of the autonomous driving function model based on the first output data and the constraint conditions; an input data determination module, used to determine the first input data corresponding to the autonomous driving function model based on the first output data in response to the defect detection result indicating that the autonomous driving function model has a defect; and a model repair module, used to repair the defects of the autonomous driving function model based on the first input data and the constraint conditions.

[0007] According to a third aspect of the present disclosure, a computer-readable storage medium is provided, which stores a computer program for executing the defect repair method of the autonomous driving function model provided in the embodiment of the first aspect.

[0008] The fourth aspect of the present disclosure provides an electronic device, which includes: a processor; a memory for storing processor executable instructions; a processor, configured to read executable instructions from the memory and execute the instructions to implement the defect repair method for the autonomous driving function model provided in the embodiment of the first aspect above.

[0009] The fifth aspect of the present disclosure provides a computer program product. When the instructions in the computer program product are executed by a processor, the defect repair method of the autonomous driving function model provided in the embodiment of the first aspect above is executed.

[0010] Based on the defect repair method of the autonomous driving function model provided by the present invention, the defect detection of the model is realized through the output data and constraints of the model, and the defect repair of the model is realized through the input data corresponding to the output data and the constraints. In this way, the defect detection and repair of the model can be completed without relying on the specific form of the defect, thereby enhancing the safety and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of a defect repair system for an autonomous driving function model provided by an exemplary embodiment of the present disclosure.

[0012] Figure 2 A flowchart of a defect repair method for an autonomous driving function model provided by an exemplary embodiment of the present disclosure.

[0013] Figure 3 A schematic diagram of the timing logic corresponding to a drone flight path planning scenario provided by an exemplary embodiment of the present disclosure.

[0014] Figure 4 A flowchart of a defect repair method for an autonomous driving function model provided by another exemplary embodiment of the present disclosure.

[0015] Figure 5 A flowchart of a defect repair method for an autonomous driving function model provided as yet another exemplary embodiment of the present disclosure.

[0016] Figure 6 A flowchart of a defect repair method for an autonomous driving function model provided as yet another exemplary embodiment of the present disclosure.

[0017] Figure 7 A schematic diagram of a mode corresponding to an autonomous driving ACC model provided for an exemplary embodiment of the present disclosure.

[0018] Figure 8 A schematic diagram of the structure of a defect repair device for an autonomous driving function model provided by an exemplary embodiment of the present disclosure.

[0019] Fig. 9 A schematic structural diagram of a defect repair device for an autonomous driving function model provided by another exemplary embodiment of the present disclosure.

[0020] Fig.10 A schematic structural diagram of a defect repair device for an autonomous driving function model provided as yet another exemplary embodiment of the present disclosure.

[0021] Fig.11 It is a structural schematic diagram of an electronic device provided by an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0022] To explain the present disclosure, example embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. It should be understood that the present disclosure is not limited to the example embodiments.

[0023] It should be noted that the relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless specifically stated otherwise.

[0024] Application Overview

[0025] In recent years, with the rapid development of autonomous driving systems, the security issues therein have attracted increasing attention. Modern autonomous driving systems are embedded with a large number of software, communication channels, sensors and other information processing units, making these systems more vulnerable to cyber attacks and physical attacks. For example, sensor spoofing attacks have become a prominent problem in autonomous driving systems. Hackers can arbitrarily manipulate sensor measurements, thereby destroying system security or driving the system to produce unsafe behaviors. At the same time, various strongly interactive pedestrians and non-motor vehicles on real roads also bring strong deception to autonomous driving systems. Current autonomous driving systems often cannot correctly identify and process these strongly interactive objects, thus causing safety issues. In addition, attackers can also disable the vehicle's braking system by obtaining communication channels.

[0026] At the same time, if these software, communication channels, sensors and other information processing units fail, the decisions made by the autonomous driving system will have unforeseen consequences.

[0027] However, current safety design technologies do not provide a feasible way for R&D personnel to automatically repair system models based on changing anti-attack requirements and unforeseen failure risks, which poses a challenge to building an autonomous driving system that can detect and repair defects.

[0028] Exemplary Systems

[0029] Figure 1 A defect repair system for an autonomous driving function model is provided in an embodiment of the present disclosure. The defect repair system 100 for an autonomous driving function model may include a collection module 101, a control processing module 102 and an execution module 103.

[0030] The acquisition module 101 may be an information acquisition device such as a vehicle-mounted camera, a laser radar, a millimeter-wave radar, a sound sensor, a wheel speed sensor, a steering angle sensor, etc., for collecting various environmental data inside and outside the vehicle. The control processing module 102 may be a control unit such as a CPU (Central Processing Unit), an ECU (Electronic Control Unit), etc., for processing various environmental data collected by the acquisition module 101, extracting features for intelligent driving; and performing decision planning on the extracted intelligent driving features to determine the decision planning information. The execution module 103 may be an actuator, for executing the control instructions corresponding to the decision planning information, such as accelerating, decelerating, braking and steering according to the control instructions. The actuator may be a drive motor, or an execution unit such as a steering gear. Among them, the automatic driving function model in the embodiment of the present disclosure may be deployed in the control processing module 102, whose input is various environmental data collected by the acquisition module 101, and whose output is the features of intelligent driving.

[0031] For example, Figure 1 As shown, the control processing module 102 can be used to perform the following method steps: determine the constraint conditions corresponding to the autonomous driving function model and the first output data of the autonomous driving function model; determine the defect detection result of the autonomous driving function model based on the first output data and the constraint conditions; in response to the defect detection result that the autonomous driving function model has defects, determine the first input data corresponding to the autonomous driving function model based on the first output data; and repair the defects of the autonomous driving function model based on the first input data and the constraint conditions. Thus, the defect detection of the model is realized through the output data and constraint conditions of the model, and the defect repair of the model is realized through the input data corresponding to the output data and the constraint conditions; in this way, the defect detection and repair of the model can be completed without relying on the specific form of the defect, thereby enhancing the safety and reliability of the model.

[0032] Exemplary Methods

[0033] Figure 2 The flowchart of the defect repair method of the autonomous driving function model provided by an exemplary embodiment of the present disclosure is as follows. The defect repair method of this embodiment can be applied to electronic devices, which may include but are not limited to personal computers, server computers, multi-processor systems, etc. Figure 2 As shown, the method includes steps S201 to S204.

[0034] Step S201: determine the constraint conditions corresponding to the autonomous driving function model and the first output data of the autonomous driving function model.

[0035] Exemplarily, the autonomous driving function model includes any function model corresponding to the autonomous driving system. For example, the autonomous driving function model is an autonomous driving simulation learning model. For another example, the autonomous driving function model is an autonomous driving object detection and recognition model. For another example, the autonomous driving function model is an autonomous driving object depth prediction model. That is, the embodiments of the present disclosure do not limit the specific type of the autonomous driving function model, and the model can be a machine learning model based on a neural network model, or a non-machine learning model, such as a traditional adaptive cruise control model.

[0036] In some examples, the input data of the autonomous driving function model includes environmental input and / or control input. For example, environmental input may be any combination of one or more of road information, high-definition maps, perception information, LIDAR information, and millimeter-wave radar information. Control input may be any combination of one or more of speed information, steering wheel information, etc. The output of the autonomous driving function model may be one or more of trajectory, object detection and recognition results, and the amount of vehicle crossing the line. In addition, during the driving of the vehicle, the autonomous driving function model will generate output data at each moment, and the first output data is the output data of the autonomous driving function model at each moment.

[0037] In some examples, the constraints corresponding to the autonomous driving function model may be temporal logic constraints, which are formal specification languages ​​that describe the behavior of time series data. It combines logical expressions and time constraints, and can be used to specify the desired system behavior and verify the behavior of the system. That is, temporal logic can be used to describe safety constraints and goals on time series data, which allows users to define various properties about the signal, such as duration, range of values, rate of change, etc., and express these properties in a formalized way. It should be noted that the autonomous driving function model itself also has some safety constraints, which are one or more constraints based on mathematical expressions, so that the model meets safety and other performance requirements. These safety constraints of the model itself are part of the constraints corresponding to the autonomous driving function model in the embodiment of the present disclosure, and the constraints in the embodiment of the present disclosure also include the system behavior corresponding to the autonomous driving function model.

[0038] Figure 3 A schematic diagram of the timing logic corresponding to the UAV flight route planning scenario, such as Figure 3 As shown, the horizontal and vertical axes are the horizontal and vertical coordinates. Within the time range of 0-T, the drone 31 needs to avoid the two dangerous areas, dangerous area 1 and dangerous area 2, and finally reach the desired target area 3. Therefore, the timing logic of the drone can be described by the following formula (1):

[0039]

[0040] Among them, [0,T] is the time constraint, ¬ is the symbol for "not", □ is the symbol for indicating the duration, Unsafe 1 represents the dangerous area 1, Unsafe 2 represents the dangerous area 2.

[0041] Of course, the temporal logic of the above formula (1) is the simplest example. In the actual application process, the temporal logic constraint conditions of the UAV can also be set from aspects such as target constraints, path constraints, and time constraints based on the speed, position, wind speed, etc. of the UAV. Similarly, for the autonomous driving function model, the constraint conditions of the model can also be determined from aspects such as target constraints, path constraints, time constraints, and safety constraints. Moreover, the constraint conditions corresponding to different autonomous driving function models are different; the constraint conditions corresponding to the same autonomous driving function model under different systems can be the same or different.

[0042] In some embodiments, when the autonomous driving function model is an autonomous driving imitation learning model, the constraint conditions can be determined in the following way: determining the model type of the autonomous driving imitation learning model and the system task corresponding to the autonomous driving imitation learning model; based on the model type of the autonomous driving imitation learning model and the system task corresponding to the autonomous driving learning model, determining the safety constraint and reachability constraint corresponding to the output trajectory of the autonomous driving imitation learning model; based on the safety constraint and reachability constraint corresponding to the output trajectory, determining the constraint conditions corresponding to the autonomous driving imitation learning model.

[0043] Exemplarily, the constraint conditions corresponding to the autonomous driving function model can be determined through the model type of the autonomous driving function model and the system task of the system where the autonomous driving function model is located. Among them, the model type of the autonomous driving function model includes but is not limited to: object recognition model, object detection model, image segmentation model, depth estimation model, trajectory prediction model, etc. The system task of the system where the autonomous driving function model is located refers to a series of operations or activities carried out in the autonomous driving system to achieve a predetermined goal, that is, a process of achieving the goal. For example, Figure 3The system task in refers to the expected system behavior, that is, it is necessary to bypass the dangerous area 1 and the dangerous area 2 to reach the target area 3 within the duration of 0-T. Therefore, the most basic safety constraints of the model itself can be determined by the model type of the autonomous driving function model. For example, the speed limit of the vehicle during autonomous driving. At the same time, the safety constraints of the task can be determined by the system task of the system where the autonomous driving function model is located. Furthermore, the most basic safety constraints of the model and the safety constraints of the task together constitute the safety constraints in the constraint conditions of the autonomous driving function model. The outputs of different model types are different, so the constraints can also include other constraints in addition to safety constraints. For example, if the autonomous driving function model is an autonomous driving imitation learning model, and the output of the model is the future driving trajectory of the car, the constraints can also include the accessibility constraints corresponding to the output trajectory. In addition, since the driving process of the vehicle is a continuous process related to time, the constraints in the embodiment of the present disclosure can be related to time, that is, the constraints are temporal logic constraints. For example, the temporal logic constraints of the autonomous driving function model are that when driving from A to B, you need to stop when you encounter a traffic light, you need to slow down when the distance between vehicles is less than a preset value, and the trajectory limit and speed limit during driving, etc. The corresponding constraints for different autonomous driving function models can be set according to the actual application, and examples are not given here one by one.

[0044] Step S202: Determine a defect detection result of the autonomous driving function model based on the first output data and the constraint conditions.

[0045] Exemplarily, if the first output data satisfies the constraint condition, it is considered that the autonomous driving function model has no defects. Conversely, if the first output data does not satisfy the constraint condition, it is considered that the autonomous driving function model has defects. It can be seen that the technical means used to detect model defects in the embodiments of the present disclosure are not limited to the specific form of defects, and both external attacks and faults of the vehicle system model itself can be detected.

[0046] In some examples, the first output data includes: one or more of the trajectory of the vehicle, the distance between the vehicle and the object, the control command of the vehicle, the amount of the vehicle's line pressure, the detection and recognition results of the object, etc. For example, if the autonomous driving function model is an autonomous driving imitation learning model, the first output data is the driving trajectory of the vehicle in the future. The constraints include: one or more of the target constraints, safety constraints, accessibility constraints, etc. Among them, the specific contents of the target constraints, safety constraints and accessibility constraints can be set according to the specific circumstances of the autonomous driving function model actually used. For example, if the autonomous driving function model is an autonomous driving imitation learning model, the constraints include safety constraints and accessibility constraints. Among them, the safety constraints include the distance between the vehicle and the vehicle in front, the collision avoidance performance of the vehicle, whether the vehicle will cross the curb, etc., and the accessibility constraints include whether the vehicle can follow the navigation and whether the vehicle can reach the set destination.

[0047] It should be noted that during the driving process of the vehicle, the output data of the autonomous driving function model will be detected in real time to determine whether the vehicle is attacked by the outside world at a certain moment or whether the vehicle itself has a fault, so as to ensure the safety of the vehicle.

[0048] In some examples, Figure 1 The control processing module 102 in can include a control module, a supervision module, etc., and then the above step S202 can be implemented by the supervision module: the supervision module can check whether the output of the autonomous driving function model meets expectations according to the definition of constraints (such as sequential logic constraints). If it does not meet expectations, the model is repaired. That is, the supervision module will continuously monitor whether the system corresponding to the model continues to safely achieve the task objectives during operation. The function of the supervision module is to capture the erroneous behavior of the system model with the assistance of constraints within a limited time range to check whether there are faults and attacks on the erroneous behavior of the system model for the correctness attribute.

[0049] In some examples, the autonomous driving function model may be an autonomous driving imitation learning model, the input data of which is map information based on GPS signals and perception information of the vehicle's surrounding environment (for example, the vehicle's own camera captures real-time environmental information). The input data is injected into the autonomous driving imitation learning model (the autonomous driving imitation learning model can be implemented end-to-end based on a deep neural network) to obtain corresponding output data, for example, the output data is the vehicle's driving trajectory for the next 6 seconds. Then, the output data of the model is input into the sequential logic detection module. The sequential logic detection module detects whether the output trajectory meets safety constraints (such as the distance between the vehicle and the vehicle in front, the vehicle's anti-collision performance, whether the vehicle will cross the curb, etc.) and accessibility constraints (such as whether the vehicle can follow the navigation, whether the vehicle can reach the set destination, etc.), so as to determine whether the autonomous driving imitation learning model has defects.

[0050] Step S203: In response to the defect detection result indicating that the autonomous driving function model has defects, determine the first input data corresponding to the autonomous driving function model based on the first output data.

[0051] Exemplarily, the first input data includes at least environmental input and / or control input. For example, environmental input may be any combination of one or more of road information, high-definition map, perception information, LIDAR information, and millimeter-wave radar information. Control input may be any combination of one or more of speed information, steering wheel information, etc. The first input data is processed using the autonomous driving function model, and the output data of the autonomous driving function model is the first output data. In an embodiment of the present application, if it is determined based on the first output data and the constraint condition that the autonomous driving function model has defects, it is necessary to determine the first input data corresponding to the first output data in turn. For example, if the autonomous driving function model is an autonomous driving imitation learning model, the first output data of the model is the trajectory of the vehicle in the next 6 seconds, and based on the trajectory and the corresponding constraint condition, it is determined that the autonomous driving imitation learning model has defects, then it is necessary to determine the model input corresponding to the first output data (such as the specific content of the input map information and perception information), and determine the specific content of the map information and perception information as the first input data.

[0052] In some examples, a target data set may be obtained first, and then based on the first output data, the first input data corresponding to the autonomous driving function model may be determined from the target data set by traversal search. If the autonomous driving function model is in an offline stage (i.e., a training stage), the target data set is a training data sample set. If the autonomous driving function model is in an online stage (i.e., a use stage), all input data of the autonomous driving function model at the previous moment are first determined, and the collection of all input data and the training data sample set is determined as the target data set.

[0053] It should be noted that, in addition to traversal search, the first input data can also be determined in other ways in the embodiment of the present disclosure, and the embodiment of the present disclosure does not limit this. In addition, the specific content of the target data set in the embodiment of the present disclosure can be determined according to actual applications, and the above is only an exemplary description.

[0054] Step S204: repair defects in the autonomous driving function model based on the first input data and constraint conditions.

[0055] Exemplarily, after determining that the autonomous driving function model has defects, the defects of the autonomous driving function model can be repaired based on the first input data and the constraints corresponding to the first output data that do not meet the constraints. For example, based on the input data corresponding to the first output data that does not meet the constraints, the model parameters of the autonomous driving function model can be updated (such as updating the weights of the model), the model structure of the autonomous driving function model can be modified (such as adding a dropout layer), and the loss function of the autonomous driving function model can be modified (such as adding a regularization term).

[0056] The defect repair method of the autonomous driving function model provided by the embodiment of the present disclosure realizes model defect detection through the output data and constraints of the model, and realizes model defect repair through the input data corresponding to the output data and the constraints. In this way, the defect detection and defect repair of the model can be completed without relying on the specific form of the defect, thereby enhancing the safety and reliability of the model.

[0057] like Figure 4 As shown in the above Figure 2 Based on the illustrated embodiment, step S204 may include steps S2041 to S2043.

[0058] Step S2041: Determine the defect type of the autonomous driving function model based on the first input data and / or constraint conditions.

[0059] Exemplarily, based on the first input data and / or constraint conditions whose output does not meet the constraint conditions, the defect type of the autonomous driving function model can be determined, and the defect type includes but is not limited to: model defects caused by data problems, defects of the model itself. For example, if there is no other input data similar to the first input data in the target data set (such as the similarity is less than the preset threshold), it is considered that the model defect is caused by data problems, such as it may be caused by GPS (Global Positioning System, Global Positioning System) positioning error or camera imaging deviation, the GPS positioning error may be caused by the attacker arbitrarily changing the GPS measurement value, and the camera imaging deviation may be caused by camera failure. For another example, if there is at least one other input data similar to the first input data in the target data set (such as the similarity is greater than or equal to the preset threshold), and the model output data corresponding to the similar other input data meets the constraint conditions, it is considered to be a defect of the model itself. Furthermore, different repair strategies can be used to repair the model defects of the autonomous driving function model for different types of defects in the autonomous driving function model.

[0060] Step S2042: Determine a target repair strategy for the autonomous driving function model based on the defect type of the autonomous driving function model.

[0061] Exemplarily, the defect types of the autonomous driving function model include but are not limited to: model defects caused by data problems and defects in the model itself. The repair strategy of the autonomous driving function model includes but is not limited to: updating the model parameters of the autonomous driving function model, updating the model structure of the autonomous driving function model, and updating the loss function corresponding to the autonomous driving function model. Furthermore, based on the specific defect type of the autonomous driving function model, the corresponding target repair strategy can be determined.

[0062] In some examples, different defects of the autonomous driving function model correspond to different repair strategies. For example, if the model defect is caused by a data problem, the corresponding repair strategy may be to retrain the autonomous driving function model using other data except the first input data to update the model parameters of the autonomous driving function model until all the outputs of the model meet the constraints; in this way, the influence of the first input data on the model during the previous model training process can be repaired. For another example, if it is a defect of the model itself, the corresponding repair strategy may be to adjust the model structure of the autonomous driving function model and / or adjust the loss function of the autonomous driving function model based on the first input data; and then retrain the adjusted model with the input data in the target data set until all the outputs of the model meet the constraints. For another example, if it is a defect of the model itself, the corresponding repair strategy may also be to retrain the autonomous driving function model based on the first input data until the output of the model meets the constraints. Furthermore, after determining the specific defect type of the autonomous driving function model, the repair strategy corresponding to the specific defect type may be selected as the target repair strategy to repair the autonomous driving function model.

[0063] Step S2043: repair the defects of the autonomous driving function model based on the target repair strategy.

[0064] Exemplarily, after determining the target repair strategy, the autonomous driving function model may be updated based on the target repair strategy to repair defects in the autonomous driving function model.

[0065] The defect repair method of the autonomous driving function model provided in the embodiment of the present disclosure determines the specific defect type of the model through the first input data and / or constraint conditions of the model, and then selects the corresponding target repair strategy to repair the model. In this way, the defect repair work of the model can be completed accurately without relying on the specific form of the defect, thereby improving the accuracy of model defect repair.

[0066] like Figure 5 As shown in the above Figure 4 Based on the illustrated embodiment, step S2041 may include steps S21a to S23a.

[0067] Step S21a, determining a target data set and a preset threshold; wherein the target data set includes first input data and a plurality of second input data.

[0068] Exemplarily, if the autonomous driving function model is in an offline stage (i.e., a training stage), the target data set includes a training data sample set. If the autonomous driving function model is in an online stage (i.e., an actual use stage), the target data set includes the input data of the autonomous driving function model at a previous moment and a training data sample set. Furthermore, the target data set includes a first input data and a plurality of second input data, and the first input data and each second input data are different input data.

[0069] It should be noted that the specific content included in the target data set can be set according to actual usage, and the embodiments of the present disclosure do not limit this.

[0070] Step S22a: compare each second input data with the first input data to obtain a comparison value corresponding to each second input data.

[0071] Exemplarily, each second input data in the plurality of second input data is respectively compared with the first input data for similarity, and a similarity comparison value corresponding to each second input data is obtained. For example, taking the target data set including the second input data A, the second input data B and the second input data C as an example, the second input data A is compared with the first input data for similarity to obtain a comparison value 1, the second input data B is compared with the first input data for similarity to obtain a comparison value 2, and the second input data C is compared with the first input data for similarity to obtain a comparison value 3.

[0072] In some examples, the KL divergence (Kullback-Leibler Divergence) algorithm can be used to compare the similarity between each second input data and the first input data. Of course, other similarity comparison methods can also be used to determine the similarity between each second input data and the first input data, and the embodiments of the present disclosure are not limited to this.

[0073] Step S23a, determining the size relationship between the comparison value corresponding to each second input data and the preset threshold, and determining the defect type of the autonomous driving function model based on the size relationship and / or constraint conditions.

[0074] Exemplarily, the size relationship between the comparison value corresponding to each second input data and the preset threshold value can be determined first, and then the defect type of the autonomous driving function model can be determined based on the size relationship, or the defect type of the autonomous driving function model can be determined based on the size relationship and the constraint condition. Among them, the preset threshold value can be set according to the actual usage. For example, if the comparison values ​​corresponding to each second input data are all less than the preset threshold value, it is determined that the autonomous driving function model is a first type of defect. If there is a comparison value greater than or equal to the preset threshold value among the comparison values ​​corresponding to each second input data, then the defect type of the autonomous driving function model is determined to be the first type or the second type based on the constraint condition.

[0075] The defect repair method of the autonomous driving function model provided by the embodiment of the present disclosure determines the defect type of the autonomous driving function model by determining whether there is second input data similar to the first input data whose output does not meet the constraint conditions from multiple second input data, and the constraint conditions corresponding to the model; in this way, the defect detection of the model can be completed in a result-oriented manner without relying on the specific form of the defect, and the specific type of the defect can be accurately determined, so that the autonomous driving function model can be repaired accordingly according to different situations.

[0076] In some embodiments, the defect type of the autonomous driving function model is determined based on size relationships and / or constraints, including: in response to the size relationship being that there is at least one comparison value greater than or equal to a preset threshold, determining at least one target input data corresponding to the at least one comparison value; using the autonomous driving function model to process the at least one target input data to obtain at least one target output data corresponding to the autonomous driving function model; in response to at least one target output data not satisfying the constraint conditions, determining that the defect type of the autonomous driving function model is a first type; in response to the presence of target output data satisfying the constraint conditions in at least one target output data, determining that the defect type of the autonomous driving function model is a second type.

[0077] Exemplarily, if the size relationship indicates that there is at least one comparison value greater than or equal to a preset threshold, at least one second input data corresponding to the at least one comparison value greater than or equal to the preset threshold is determined, and the at least one second input data is determined as the target input data. Further, at least one target output data corresponding to the at least one target input data is determined. If none of the at least one target output data satisfies the constraint condition, the defect type of the autonomous driving function model is determined to be a model defect caused by a data problem, that is, the first input data and the at least one target input data may be erroneous data or inaccurate data caused by reasons such as GPS positioning errors or camera imaging deviations. At this time, the defect type of the autonomous driving function model is determined to be the first type. If there is a target output data that satisfies the constraint condition in the at least one target output data, the defect type of the autonomous driving function model is determined to be a model defect rather than a problem with the input data. At this time, the defect type of the autonomous driving function model is determined to be the second type.

[0078] In some embodiments, the defect type of the autonomous driving function model is determined based on size relationships and / or constraints, including: in response to the size relationship being that the comparison values ​​corresponding to each second input data are all less than a preset threshold, determining that the defect type of the autonomous driving function model is the first type.

[0079] Exemplarily, if the size relationship indicates that the comparison value corresponding to each second input data in multiple second input data is less than a preset threshold, then the defect type of the autonomous driving function model is determined to be a model defect caused by data problems, that is, the first input data may be erroneous data or inaccurate data due to reasons such as GPS positioning errors or camera imaging deviations. At this time, the defect type of the autonomous driving function model is determined to be the first type.

[0080] like Figure 6 As shown in the above Figure 4 Based on the illustrated embodiment, step S2042 may include steps S21b to S22b.

[0081] Step S21b: In response to the defect type of the autonomous driving function model being the first type, determining that a target repair strategy for the autonomous driving function model is to update model parameters of the autonomous driving function model.

[0082] For example, if the defect type of the autonomous driving function model is the first type, the first type is a model defect caused by a data problem. At this time, the repair strategy corresponding to the autonomous driving function model is to update the model parameters of the autonomous driving function model. For example, the autonomous driving function model can be retrained with accurate input data to update the model's weights, learning rate and other parameter information.

[0083] Step S22b, in response to the defect type of the autonomous driving function model being the second type, determining that a target repair strategy for the autonomous driving function model is to update the model structure of the autonomous driving function model and / or update the loss function of the autonomous driving function model.

[0084] Exemplarily, if the defect type of the autonomous driving function model is the second type, the second type is a defect of the model itself. At this time, the repair strategy corresponding to the autonomous driving function model is to update the model structure of the autonomous driving function model, or to update the loss function of the autonomous driving function model, or to update the model structure and loss function of the autonomous driving function model. Then, the autonomous driving function model with the updated model structure and / or loss function is retrained using multiple input data including the first input data. Among them, updating the model structure of the autonomous driving function model includes but is not limited to: adjusting the number of hidden layer nodes, increasing the number of hidden layers, adjusting the activation function, adjusting the batch size of the input data, etc.

[0085] It should be noted that, regardless of whether it is the repair strategy corresponding to the first type of defect or the repair strategy corresponding to the second type of defect, the output of the model must meet the constraint conditions when the model is retrained, and the model is considered to be retrained.

[0086] In some examples, when the autonomous driving function model is an autonomous driving imitation learning model, if the sequential logic detection module detects that the first output trajectory of the model does not meet the constraints (safety constraints and reachability constraints), the first output trajectory is injected into the pre-processing module (the pre-processing module is located at the front end of the autonomous driving imitation learning model), and the first input data corresponding to the first output trajectory is found through traversal search. There are two situations at this time: (1) If there are multiple other input data that are highly similar to the first input data, and at least one output data in the model output data corresponding to the other input data meets the constraints, then the reason can be located that the generalization ability of the model is not good. At this time, the model structure (such as adding a dropout layer) or the loss function (such as adding a regularization term) is updated, and then the multiple input data including the first input data are injected into the updated model to obtain a new output, and the model conversion is completed. (2) If there are no other input data that are highly similar to the first input data, the reason can be located that it is due to data acquisition errors or data attacks. At this time, the first input data is removed, and the original model is trained based on accurate input data or input data that has not been attacked to obtain a new output trajectory, and the model conversion is completed. Of course, the model conversion methods in (1) and (2) above need to be executed until all model outputs meet the constraints.

[0087] The defect repair method of the autonomous driving function model provided by the embodiment of the present disclosure repairs the model by adopting corresponding target repair strategies for different types of defects, and repairs the model from multiple angles such as model defects caused by data problems and the generalization ability of the model itself. In this way, the type of defect can be determined without relying on the specific form of the defect, so as to repair the model according to different situations, thereby improving the defect repair efficiency and defect repair effect of the model.

[0088] In some embodiments, when it is determined that the defect type of the autonomous driving function model is of the first type, the defect of the autonomous driving function model is repaired based on a target repair strategy, including: determining a target data set; using other input data in the target data set except the first input data to train the autonomous driving function model until all output data of the autonomous driving function model meet the constraint conditions.

[0089] Exemplarily, when the defect type of the autonomous driving function model is the first type, the input data corresponding to the model output that does not meet the constraint conditions can be eliminated, and the model can be retrained using data that meets the requirements, thereby preventing the model from being affected by inaccurate data or attacked data used in previous training.

[0090] In some embodiments, when it is determined that the defect type of the autonomous driving function model is the second type, the defect of the autonomous driving function model is repaired based on the target repair strategy, including: updating the model structure and / or loss function of the autonomous driving function model to obtain an updated autonomous driving function model and / or an updated loss function; using the updated autonomous driving function model and / or the updated loss function to process the first input data until the output of the updated autonomous driving model meets the constraint conditions.

[0091] For example, when the defect type of the autonomous driving function model is the second type, the structure of the model and / or the loss function corresponding to the model may be modified. Then, the modified model and / or the loss function are retrained using a data set including the first input data and / or input data highly similar to the first input data until all outputs of the model meet the constraint conditions.

[0092] In some embodiments, when the autonomous driving function model is an autonomous driving imitation learning model and the first output data is a first output trajectory, a defect detection result of the autonomous driving function model is determined based on the first output data and constraints, including: in response to the first output trajectory satisfying the safety constraints and reachability constraints in the constraints, determining that the autonomous driving imitation learning model has no defects; in response to the first output trajectory not satisfying the safety constraints or the first output trajectory not satisfying the reachability constraints, determining that the autonomous driving imitation learning model has defects.

[0093] Exemplarily, if the first output trajectory of the autonomous driving imitation learning model satisfies both the safety constraint and the reachability constraint, the autonomous driving imitation learning model is considered to have no defects. If the first output trajectory of the autonomous driving imitation learning model does not satisfy the safety constraint, or the first output trajectory of the autonomous driving imitation learning model does not satisfy the reachability constraint, the autonomous driving imitation learning model is considered to have defects.

[0094] In some embodiments, model defects can be repaired by model conversion. That is, a new model that meets expectations can be generated based on the original model, the first input data and the constraint conditions to complete the model repair. The detailed process can be described as follows:

[0095] 1) For the original model m 0 Make a copy to get the new model m 1 .

[0096] 2) Determine the parameter value that causes the model to violate the constraint conditions through parameter search, and find the corresponding negative examples (i.e., the input data corresponding to the model output that does not meet the constraint conditions, such as the first input data) for guidance. 1The mode where the parameter value that violates the constraint condition is converted, and the new model m is updated in the process 1 The corresponding parameters in . Among them, mode refers to the mode state of the model, and the model can be in different modes at different times.

[0097] 3) Continue to determine the new model m based on the constraints 1 The output of the new model m is considered to meet expectations, and the above step 2) is repeated to transform the modes that do not meet expectations and update the corresponding parameters. 1 When all outputs meet expectations, the model conversion is considered complete.

[0098] Figure 7 A schematic diagram of a mode corresponding to an automatic driving ACC model provided in an embodiment of the present disclosure, such as Figure 7 As shown in FIG. 1 , a simple automatic driving ACC (Adaptive Cruise Control) system model realizes the normal operation of the automatic driving system through speed control mode 71 and distance control mode 72. In speed control mode 71, the vehicle has two state variables: the distance d from the vehicle to the vehicle in front and the speed v of the vehicle. The ACC system is equipped with two sensors: one sensor measures the speed of the vehicle through a noisy wheel encoder, that is, v wheel =v+n wheel ; A sensor measures the speed of the ego vehicle through a noisy GPS sensor, namely v gps =v+n gps ; where n wheel and n gps Represent the noise of the wheel encoder and the noise of the GPS respectively. At the same time, the ACC system measures the distance through a radar sensor, namely d r =d+n r , n r Represents the noise of the radar sensor. The ACC system decides to use speed control mode 71 or distance control mode 72 based on the real-time sensor measurement. The ACC system needs to switch modes according to the distance between the ego vehicle and the other vehicle. If the ego vehicle is too close to the vehicle in front, the controller in the ACC system will trigger a mode change, so that the distance between the ego vehicle and the vehicle in front increases, and the ACC system will switch from speed control to distance control. Similarly, if the ego vehicle is farther away from the vehicle in front, the ACC system will switch from distance control to speed control, and maintain safe speed and safe distance control by executing this transition.

[0099] However, GPS spoofing is a serious problem in the actual implementation of ACC systems. GPS spoofing refers to the GPS receiver receiving erroneous GPS data packets, which allows attackers to arbitrarily change GPS speed measurements, thus causing safety incidents.

[0100] In view of the above-mentioned problems of the ACC model of autonomous driving, the output data of the ACC system model can be monitored in real time. If it is found that the output data of the ACC system model at a certain moment does not meet the constraints (such as the output of the model does not meet the measured value of the distance to the front vehicle or the noise distance), it is determined that the ACC system model has defects. At this time, the input data of the ACC system model at this moment is determined, and the ACC system model is repaired based on the input data and constraints at this moment. For example, in the target data set, it is determined whether there are other input data that are highly similar to the input data at this moment. If there are no other input data that are highly similar to the input data at this moment, or there are other input data that are highly similar to the input data at this moment, but the model output corresponding to the other highly similar input data also does not meet the constraints, it is determined that the model defect is caused by data problems at this moment. Then, the input data at this moment can be analyzed, and the analysis result obtained is that there is GPS spoofing in the input data at this moment. Finally, the GPS value can be ignored by switching the mode, and the speed can be estimated only by using the wheel encoder. That is, a copy of the original model is first created, in which the controller in the ACC system simply ignores the GPS reading because it is no longer credible. Then, based on the temporal logic constraints, new transformations are added to the mode of the original model and the corresponding replica instances, which only use wheel encoders to measure speed. Thus, the relevant parameters are updated through mode transformation, which specify the threshold for the ACC model to switch from the original replica to the new replica to meet safety requirements and achieve model repair. Among them, the temporal logic constraints corresponding to the ACC system model include: the limit of GPS noise threshold, the distance between the vehicle and the front vehicle, etc.

[0101] In some embodiments, if the model is a traditional non-machine learning model, the model repair process is the process of determining new model parameters. Since the structure of the model has been determined, the model repair problem is the problem of generating model parameters. Therefore, P can be defined first. 1 For the new model m 1 The parameter set, given the constraints and parameter set P 0 , the optimization objective can be converted into 0 Find a set of parameters P 1 , so that The detailed steps of the process are as follows:

[0102] 1) Define the initial state variables and parameters of the model, which can be obtained from the original model m in the model conversion 0 At the same time, define a set of parameters P that need to be optimized 1 and parameter set P 0 In addition, the maximum solution time of the parametric solver needs to be defined.

[0103] 2) Call the supervision module's counter-proof loop to search for negative examples based on the definition of temporal logic. In each search, if a negative example is detected, the corresponding parameter P will be returned. 1 According to these parameters, the system will perform corresponding parameter updates based on the parameter solver, namely

[0104] 3) Repeat the above step 2) until all parameter updates are completed.

[0105] 4) Finally, for all updated parameters P 1 , and combine them into the optimal parameter set If the supervision module verifies the newly generated optimal parameter set If the definition of the constraints can be met, a new model is output, the model generation process is completed, and the model repair is completed.

[0106] Exemplarily, the model conversion problem can be transformed into a parameter optimization problem of estimating a sparse parameter vector, so that parameter solution is performed to determine new parameters of the model.

[0107] In some embodiments, in addition to determining the defect detection result of the autonomous driving function model based on the first output data and constraint conditions of the model, the defect detection result of the autonomous driving function model can also be determined by using Figure 1 The control processing module 102 (such as the control module) in the defect repair system 100 processes the first output data of the model to obtain the decision planning information corresponding to the first output data, and then determines the defect detection result of the autonomous driving function model based on the decision planning information and the constraints. Of course, the defect detection result of the autonomous driving function model can also be determined based on the control commands and constraints corresponding to the decision planning information. For example, when the first output data of the model is trajectory information, the defect detection result of the autonomous driving function model can be determined using the control commands and constraints corresponding to the trajectory information. Among them, the control commands corresponding to the trajectory information include but are not limited to: vehicle distance control command, vehicle speed control command, steering wheel angle control command, steering wheel speed control command, etc.

[0108] Exemplary Devices

[0109] Figure 8 A defect repair device for an autonomous driving function model provided in an embodiment of the present disclosure, such as Figure 8 As shown, the defect repairing device 800 includes an output data determining module 801 , a model defect detecting module 802 , an input data determining module 803 and a model repairing module 804 .

[0110] An output data determination module 801 is used to determine the constraint conditions corresponding to the autonomous driving function model and the first output data of the autonomous driving function model;

[0111] A model defect detection module 802, configured to determine a defect detection result of the autonomous driving function model based on the first output data and the constraint condition;

[0112] An input data determination module 803 is used to determine, in response to a defect detection result indicating that the autonomous driving function model has a defect, first input data corresponding to the autonomous driving function model based on the first output data;

[0113] The model repair module 804 is used to repair defects in the autonomous driving function model based on the first input data and constraint conditions.

[0114] In some embodiments, Fig. 9 As shown, the model repair module 804 includes a defect judgment unit 8041 , a strategy determination unit 8042 and a model repair unit 8043 .

[0115] A defect judgment unit 8041 is used to determine a defect type of the automatic driving function model based on the first input data and / or the constraint condition;

[0116] A strategy determination unit 8042 is used to determine a target repair strategy for the autonomous driving function model based on the defect type of the autonomous driving function model;

[0117] The model repair unit 8043 is used to repair defects in the autonomous driving function model based on a target repair strategy.

[0118] In some embodiments, the defect judgment unit 8041 is used to determine a target data set and a preset threshold; wherein the target data set includes a first input data and multiple second input data; each second input data is compared with the first input data to obtain a comparison value corresponding to each second input data; the size relationship between the comparison value corresponding to each second input data and the preset threshold is determined, and the defect type of the autonomous driving function model is determined based on the size relationship and / or constraint conditions.

[0119] In some embodiments, the defect judgment unit 8041 is used to determine a target data set and a preset threshold; wherein the target data set includes a first input data and multiple second input data; each second input data is compared with the first input data to obtain a comparison value corresponding to each second input data; the size relationship between the comparison value corresponding to each second input data and the preset threshold is determined; in response to the size relationship being that there is at least one comparison value greater than or equal to the preset threshold, at least one target input data corresponding to the at least one comparison value is determined; at least one target input data is processed using the autonomous driving function model to obtain at least one target output data corresponding to the autonomous driving function model; in response to at least one target output data not satisfying the constraint conditions, the defect type of the autonomous driving function model is determined to be the first type; in response to the presence of target output data satisfying the constraint conditions in at least one target output data, the defect type of the autonomous driving function model is determined to be the second type.

[0120] In some embodiments, the defect judgment unit 8041 is used to determine a target data set and a preset threshold; wherein the target data set includes a first input data and multiple second input data; each second input data is compared with the first input data to obtain a comparison value corresponding to each second input data; the size relationship between the comparison value corresponding to each second input data and the preset threshold is determined; in response to the size relationship being that the comparison values ​​corresponding to each second input data are all less than the preset threshold, determining that the defect type of the autonomous driving function model is the first type.

[0121] In some embodiments, Fig.10 As shown, the policy determination unit 8042 includes a first policy determination subunit 81 and a second policy determination subunit 82 .

[0122] a first strategy determination subunit 81, configured to determine, in response to the defect type of the autonomous driving function model being the first type, that a target repair strategy for the autonomous driving function model is to update a model parameter of the autonomous driving function model;

[0123] The second strategy determination subunit 82 is used to determine, in response to the defect type of the autonomous driving function model being the second type, that a target repair strategy for the autonomous driving function model is to update the model structure of the autonomous driving function model and / or to update the loss function of the autonomous driving function model.

[0124] In some embodiments, when the autonomous driving function model is an autonomous driving imitation learning model and the first output data is a first output trajectory, the model defect detection module 802 is used to determine that the autonomous driving imitation learning model has no defects in response to the first output trajectory satisfying the safety constraint and the reachability constraint in the constraint conditions; and determine that the autonomous driving imitation learning model has defects in response to the first output trajectory not satisfying the safety constraint or the first output trajectory not satisfying the reachability constraint.

[0125] The beneficial technical effects corresponding to the exemplary embodiment of the defect repair device 800 can be found in the corresponding beneficial technical effects of the exemplary method part above, which will not be repeated here.

[0126] Exemplary Electronic Devices

[0127] Fig.11 A structural diagram of an electronic device 110 provided in an embodiment of the present disclosure includes at least one processor 111 and a memory 112 .

[0128] The processor 111 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 110 to perform desired functions.

[0129] The memory 112 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on a computer-readable storage medium, and the processor 111 may execute one or more computer program instructions to implement the defect repair method and / or other desired functions of the autonomous driving function model of the various embodiments of the present disclosure described above.

[0130] In one example, the electronic device 110 may further include: an input device 113 and an output device 114 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0131] The input device 113 may also include, for example, a keyboard, a mouse, etc.

[0132] The output device 114 can output various information to the outside, and may include, for example, a display, a speaker, a printer, a communication network and its connected remote output devices, etc.

[0133] Of course, to simplify, Fig.11 Only some of the components related to the present disclosure in the electronic device 110 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 110 may also include any other appropriate components.

[0134] Exemplary computer program products and computer-readable storage media

[0135] In addition to the above-mentioned methods and devices, embodiments of the present disclosure may also provide a computer program product, including computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the defect repair method of the autonomous driving function model of various embodiments of the present disclosure described in the above-mentioned "Exemplary Method" section.

[0136] The computer program product may be written in any combination of one or more programming languages ​​to write program code for performing the operations of the disclosed embodiments, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0137] In addition, an embodiment of the present disclosure may also be a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the steps in the defect repair method of the autonomous driving function model of various embodiments of the present disclosure described in the above-mentioned "Exemplary Method" section.

[0138] Computer readable storage media can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium is, for example, but not limited to, a system, device or device including electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0139] The basic principles of the present disclosure are described above in conjunction with specific embodiments. However, the advantages, strengths, effects, etc. mentioned in the present disclosure are only examples and not limitations, and cannot be considered as necessary for each embodiment of the present disclosure. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, rather than limitation, and the above details do not limit the present disclosure to being implemented by adopting the above specific details.

[0140] Those skilled in the art may make various changes and modifications to the present disclosure without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present disclosure claims and their equivalents, the present disclosure is also intended to include these modifications and variations.

Claims

1. A defect repair method for an autonomous driving function model, comprising: Determining a constraint condition corresponding to the autonomous driving function model and first output data of the autonomous driving function model; Determining a defect detection result of the autonomous driving function model based on the first output data and the constraint condition; In response to the defect detection result indicating that the autonomous driving function model has a defect, determining first input data corresponding to the autonomous driving function model based on the first output data; The defects of the autonomous driving function model are repaired based on the first input data and the constraint conditions.

2. The method according to claim 1, wherein: The repairing the defects of the autonomous driving function model based on the first input data and the constraint condition includes: Determining a defect type of the autonomous driving function model based on the first input data and / or the constraint condition; Determining a target repair strategy for the autonomous driving function model based on a defect type of the autonomous driving function model; Based on the target repair strategy, defects of the autonomous driving function model are repaired.

3. The method according to claim 2, wherein: The determining, based on the first input data and / or the constraint condition, a defect type of the autonomous driving function model includes: Determine a target data set and a preset threshold; wherein the target data set includes the first input data and a plurality of second input data; Compare each of the second input data with the first input data to obtain a comparison value corresponding to each of the second input data; Determine the size relationship between the comparison value corresponding to each of the second input data and the preset threshold, and determine the defect type of the autonomous driving function model based on the size relationship and / or the constraint condition.

4. The method according to claim 3, wherein: The determining the defect type of the autonomous driving function model based on the size relationship and / or the constraint condition includes: In response to the size relationship being that there is at least one comparison value greater than or equal to the preset threshold, determining at least one target input data corresponding to the at least one comparison value; Processing the at least one target input data using the autonomous driving function model to obtain at least one target output data corresponding to the autonomous driving function model; In response to the at least one target output data not satisfying the constraint condition, determining that the defect type of the autonomous driving function model is a first type; In response to the presence of target output data satisfying the constraint condition in the at least one target output data, it is determined that the defect type of the autonomous driving function model is the second type.

5. The method according to claim 3, wherein: The determining the defect type of the autonomous driving function model based on the size relationship and / or the constraint condition includes: In response to the size relationship being that the comparison values ​​corresponding to each of the second input data are all smaller than the preset threshold, it is determined that the defect type of the autonomous driving function model is the first type.

6. The method according to claim 2, wherein: The determining, based on the defect type of the autonomous driving function model, a target repair strategy of the autonomous driving function model includes: In response to the defect type of the autonomous driving function model being a first type, determining that a target repair strategy for the autonomous driving function model is to update a model parameter of the autonomous driving function model; In response to the defect type of the autonomous driving function model being the second type, determining that a target repair strategy for the autonomous driving function model is to update a model structure of the autonomous driving function model and / or to update a loss function of the autonomous driving function model.

7. The method according to any one of claims 1 to 6, wherein: When the autonomous driving function model is an autonomous driving imitation learning model and the first output data is a first output trajectory, determining a defect detection result of the autonomous driving function model based on the first output data and the constraint condition includes: In response to the first output trajectory satisfying the safety constraint and the reachability constraint in the constraint condition, determining that the autonomous driving imitation learning model has no defects; In response to the first output trajectory not satisfying the safety constraint or the first output trajectory not satisfying the reachability constraint, it is determined that the autonomous driving imitation learning model has defects.

8. A defect repair device for an autonomous driving function model, comprising: an output data determination module, configured to determine a constraint condition corresponding to the autonomous driving function model and first output data of the autonomous driving function model; a model defect detection module, configured to determine a defect detection result of the autonomous driving function model based on the first output data and the constraint condition; an input data determination module, configured to determine, in response to the defect detection result indicating that the autonomous driving function model has a defect, first input data corresponding to the autonomous driving function model based on the first output data; A model repair module is used to repair defects of the autonomous driving function model based on the first input data and the constraint conditions.

9. A computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is used to execute the defect repair method of the autonomous driving function model described in any one of claims 1 to 7.

10. An electronic device, comprising: processor; a memory for storing instructions executable by the processor; The processor is used to read the executable instructions from the memory and execute the instructions to implement the defect repair method of the autonomous driving function model described in any one of claims 1 to 7.

Images