An unmanned vehicle-oriented cloud-edge collaborative model updating method and system

Abstract

The application discloses an unmanned vehicle-oriented cloud-edge collaborative model updating method and system, and the system comprises a vehicle end and a cloud platform, wherein the vehicle end extracts corresponding inference rules from a rule library based on a currently detected abnormal event, and infers optimal scene data with the highest matching degree with the abnormal event from vehicle-end collected data based on the inference rules; the vehicle end sends the optimal scene data to the cloud platform; the cloud platform receives the optimal scene data sent by the vehicle end, and re-trains a new machine learning model for coping with the abnormal event by using the optimal scene data; the cloud platform sends the new machine learning model to the vehicle end; and the vehicle end receives the new machine learning model fed back by the cloud platform, and covers an old machine learning model for coping with the abnormal event saved by the vehicle end by using the new machine learning model. The technical scheme provided by the application can improve the accuracy and reliability of the model used by the unmanned vehicle end.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned vehicles, in particular to a cloud-edge collaborative model updating method and system for unmanned vehicles. BACKGROUND

[0002] At present, unmanned technology is a hot research. In order to realize unmanned driving at the vehicle end, one or more machine learning models need to be trained in the vehicle end, and various road conditions and unexpected situations in the running process of the vehicle are identified through the machine learning model, so as to control the safe and stable driving of the vehicle. Considering that the hardware device computing power of the vehicle end is relatively weak, the training of the machine learning model is often completed with the help of the powerful computing power of the edge cloud platform or the center cloud platform. In order to ensure higher communication efficiency, the edge cloud platform is more often used. The common steps are that the vehicle end collects driving information, sends the driving information to the edge cloud platform, the edge cloud platform trains the machine learning model using the driving information, and then the edge cloud platform issues the trained machine learning model to the vehicle end for use by the vehicle end.

[0003] However, the actual road conditions of the vehicle end are complex, and unexpected situations are changing, and the data generated by the vehicle end is also complex and diverse. The training method used in the prior art is to directly train the machine learning model using various data uploaded by the vehicle end. The actual application effect of the model trained in this way is not satisfactory. Since the useful data for actual unexpected situations only account for a small part of the total data, the recognition result of the machine learning model often has errors, causing the unmanned vehicle to be unable to correctly respond to unexpected situations, which not only affects the personal safety around the vehicle, but also may damage road facilities. SUMMARY

[0004] Therefore, the embodiments of the present application provide a cloud-edge collaborative model updating method and system for unmanned vehicles, thereby improving the accuracy and reliability of the model used by the vehicle end of the unmanned vehicle.

[0005] According to a first aspect, the embodiments of the present application provide a cloud-edge collaborative model updating method for unmanned vehicles, applied to a vehicle end, the method comprising: extracting a corresponding inference rule from a rule library based on a currently detected abnormal event, and inferring optimal scene data with the highest matching degree with the abnormal event from vehicle end collected data based on the inference rule; sending the optimal scene data to a cloud platform; receiving a new machine learning model fed back by the cloud platform, and covering an old machine learning model used to cope with the abnormal event saved by the vehicle end with the new machine learning model, the new machine learning model being a machine learning model used to cope with the abnormal event retrained by the cloud platform using the optimal scene data.

[0006] Optionally, the inference rules in the rule base are created by dividing the abnormal events and the sensor data into types of environment, job, hardware and software, and associating the abnormal events and the sensor data according to the corresponding types to generate the inference rules.

[0007] Optionally, the method further comprises: identifying the event type of the abnormal event, and extracting a target inference rule of the same type from the rule base according to the event type; obtaining sensor data associated with the target inference rule, and integrating the sensor data to obtain the optimal scene data.

[0008] Optionally, before extracting the corresponding inference rule from the rule base based on the currently detected abnormal event, the method further comprises: determining whether the old machine learning model needs to be updated based on the currently detected abnormal event; if the old machine learning model needs to be updated, sending the vehicle-end collected data to the cloud platform; if a retraining notification is received from the cloud platform, performing the step of extracting the corresponding inference rule from the rule base based on the currently detected abnormal event, wherein the retraining notification is a notification sent by the cloud platform when the source machine learning model saved in the cloud platform cannot make correct decisions after inputting the received vehicle-end collected data into the source machine learning model.

[0009] Optionally, the method further comprises: if a re-clipped model is received from the cloud platform, using the re-clipped model to replace the old machine learning model saved in the vehicle-end for coping with the abnormal event, wherein the re-clipped model is obtained by re-compressing and clipping the source machine learning model saved in the cloud platform when the source machine learning model can make correct decisions after inputting the received vehicle-end collected data into the source machine learning model.

[0010] According to a second aspect, embodiments of the present application provide a cloud-edge collaborative model updating method for unmanned vehicles, applied to a cloud platform, the method comprising: receiving optimal scene data sent by a vehicle end, the optimal scene data being data with the highest matching degree to an abnormal event detected by the vehicle end, the vehicle end extracting corresponding inference rules from a rule library based on the abnormal event and inferring the data from vehicle-end collected data based on the inference rules; retraining a new machine learning model for dealing with the abnormal event using the optimal scene data, and sending the new machine learning model to the vehicle end, so that the vehicle end covers an old machine learning model saved by the vehicle end with the new machine learning model.

[0011] Optionally, before the receiving optimal scene data sent by the vehicle end, the method further comprises: receiving vehicle-end collected data sent by the vehicle end, the vehicle-end collected data being sent by the vehicle end when the vehicle end judges that the old machine learning model needs to be updated based on the currently detected abnormal event; inputting the vehicle-end collected data into a source machine learning model saved by the cloud platform, and judging whether the source machine learning model can make correct decisions; if the source machine learning model cannot make correct decisions, sending a retraining notification to the vehicle end.

[0012] Optionally, the method further comprises: if the source machine learning model can make correct decisions, re-compressing and re-cutting the source machine learning model to obtain a re-cut model for dealing with the abnormal event, and sending the re-cut model to the vehicle end, so that when the vehicle end receives the re-cut model, the vehicle end covers the old machine learning model saved by the vehicle end with the re-cut model.

[0013] Optionally, the retraining a new machine learning model for dealing with the abnormal event using the optimal scene data comprises: retraining a new source machine learning model using the optimal scene data; compressing and cutting the new source machine learning model to obtain the new machine learning model; inputting the optimal scene data into the new machine learning model to verify whether the new machine learning model can make correct decisions; if the new machine learning model can make correct decisions, retaining the new machine learning model; if the new machine learning model cannot make correct decisions, returning to the step of compressing and cutting the new source machine learning model until the new machine learning model can make correct decisions.

[0014] According to a third aspect, embodiments of the present invention provide a cloud-edge collaborative model update system for autonomous driving. The system includes a vehicle-side device and a cloud platform, wherein: the vehicle-side device extracts corresponding inference rules from a rule base based on currently detected abnormal events, and infers the optimal scenario data with the highest matching degree to the abnormal events from data collected by the vehicle-side device based on the inference rules; the vehicle-side device sends the optimal scenario data to the cloud platform; the cloud platform receives the optimal scenario data sent by the vehicle-side device, and retrains a new machine learning model for responding to the abnormal events using the optimal scenario data; the cloud platform sends the new machine learning model to the vehicle-side device; the vehicle-side device receives the new machine learning model fed back by the cloud platform, and uses the new machine learning model to overwrite the old machine learning model stored by the vehicle-side device for responding to the abnormal events.

[0015] The technical solution provided in this application has the following advantages:

[0016] The technical solution provided in this application establishes inference rules for abnormal data collection in response to various sudden anomalies in autonomous driving. When the machine learning model on the vehicle needs to be updated, the vehicle inputs the detected abnormal event into the rule base and matches it among various preset rules. Based on the matched inference rules, abnormal data inference is performed to deduce the optimal scenario data with the highest matching degree from the data collected on the vehicle. This optimal scenario data is then sent to the cloud platform, allowing the cloud platform to retrain a new machine learning model to handle abnormal events. Finally, the vehicle receives the new machine learning model from the cloud platform and uses it to overwrite the old machine learning model stored on the vehicle for handling abnormal events. This embodiment, based on abnormal events and inference rules, can adopt appropriate data collection strategies according to different abnormal scenarios, mining effective data for specific scenarios. The new machine learning model retrained using effective data is more accurate and reliable in identifying and making decisions about abnormal events, thus significantly improving the reliability and effectiveness of model updates. Replacing the old machine learning model on the vehicle with the retrained model can reduce vehicle safety problems caused by decision-making errors on the vehicle. Attached Figure Description

[0017] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0018] Figure 1 A schematic diagram of the structure of a cloud-edge collaborative model update system for autonomous driving is shown in one embodiment of the present invention;

[0019] Figure 2 A schematic diagram of an abnormal data acquisition process is shown in one embodiment of the present invention;

[0020] Figure 3 A schematic diagram of an anomaly model structure for an engineering vehicle operation scenario is shown in one embodiment of the present invention;

[0021] Figure 4 The diagram illustrates the steps of a cloud-edge collaborative model update method for autonomous driving according to one embodiment of the present invention.

[0022] Figure 5 This diagram illustrates another step of a cloud-edge collaborative model update method for autonomous driving according to one embodiment of the present invention.

[0023] Figure 6 This diagram illustrates another step in a cloud-edge collaborative model update method for autonomous driving according to one embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] When the machine learning model on the autonomous vehicle fails to make correct decisions regarding abnormal events, existing technologies often require updating this model. The update process involves training with full datasets. The vehicle sends all data collected by its sensors to an edge cloud platform, which trains a new machine learning model. The vehicle then receives the new model returned by the edge cloud platform, thus achieving iterative model updates. However, due to the large amount of redundant and invalid data in the full dataset, the training effect is poor. To address this problem, this invention proposes the following solution.

[0026] Please see Figure 1 In one embodiment, this invention provides a cloud-edge collaborative model update system for autonomous driving, comprising a vehicle-side device 1 and a cloud platform 2, wherein:

[0027] Vehicle-side 1 extracts the corresponding inference rules from the rule base based on the currently detected abnormal events, and infers the optimal scenario data with the highest matching degree with the abnormal events from the data collected from the vehicle-side based on the inference rules;

[0028] Vehicle 1 sends the optimal scenario data to cloud platform 2;

[0029] The cloud platform 2 receives the optimal scenario data sent by the vehicle terminal 1 and uses the optimal scenario data to retrain a new machine learning model for dealing with abnormal events.

[0030] Cloud platform 2 sends the new machine learning model to vehicle terminal 1;

[0031] Vehicle 1 receives a new machine learning model from cloud platform 2 and uses the new machine learning model to overwrite the old machine learning model stored on vehicle 1 for dealing with abnormal events.

[0032] Specifically, in this embodiment, the model update system includes vehicle-side 1 and cloud platform 2. The powerful computing capabilities of cloud platform 2 are primarily used to assist in updating the machine learning model on vehicle-side 1. Cloud platform 2 includes, but is not limited to, a central cloud platform 2 and an edge cloud platform 2. In this embodiment, considering the closer communication distance between edge cloud platform 2 and vehicle-side 1, resulting in lower communication costs, edge cloud platform 2 is preferentially chosen to implement the above model update scheme. If the user has higher requirements for the computing power of model training, the central cloud platform 2 can also be chosen to implement the above scheme. This invention does not impose any special limitations on this.

[0033] In this embodiment, the vehicle-side data collection refers to the data collected by various sensors on vehicle-side 1, covering all data that can be used for anomaly detection in autonomous vehicles. It mainly includes external environmental data collected by sensor devices such as lidar, ultrasonic radar, and cameras on vehicle-side 1, as well as internal data such as instruction data of software operation in the vehicle computer and input / output data of hardware circuits.

[0034] like Figure 2As shown, when the machine learning model of vehicle-side 1 needs to be updated, the corresponding inference rules are extracted from the preset rule base based on the abnormal events detected by vehicle-side 1. These inference rules are used to infer the causes of the abnormal events and establish a mapping between the abnormal causes and the data to be collected. This allows for automatic on-demand data collection based on the specific abnormal cause. Responding to the inference rules, the vehicle-side system infers the optimal scenario data with the highest matching degree to the abnormal event from the data collected by the vehicle-side. Only this optimal scenario data is uploaded to cloud platform 2, allowing cloud platform 2 to retrain a new machine learning model to handle abnormal events using the optimal scenario data. Cloud platform 2 then distributes the trained new machine learning model to vehicle-side 1, where it replaces the old machine learning model used to handle abnormal events, completing the model update process for vehicle-side 1. Using the model update scheme provided by this system, appropriate data collection strategies can be adopted according to different abnormal scenarios, allowing for the mining of effective data for specific scenarios. The new machine learning model, retrained with effective data, is more accurate and reliable in identifying and making decisions about abnormal events. It avoids the problem of redundant data affecting training results caused by using all data to train the model, thus significantly improving the reliability and effectiveness of model updates. Replacing the old machine learning model on vehicle 1 with the retrained machine learning model can reduce vehicle safety problems caused by errors in vehicle 1's decisions.

[0035] Furthermore, in this embodiment, the process of creating the rule base and inference rules is implemented according to the type of abnormal event and the type of sensor data. Specifically, the autonomous driving scenario addressed by this invention is mainly applied to engineering machinery application scenarios; therefore, an anomaly model of the engineering vehicle operation scenario is first created to achieve situational awareness of the vehicle during operation. For example... Figure 3As shown, the created model structure includes three levels: basic categories, subcategories, and attributes. For example, "Operation" is defined as a basic category, "Behavior" as a subcategory, and "Acceleration" as an attribute. Other basic categories include at least "Environment," "Hardware," "Vehicle Software," "Cloud Platform," "Status," and "Basic Information." Each basic category further includes its own subcategories and attributes. This model structure can be seen as a classification of abnormal events, with attributes covering the smallest scope and representing the most detailed type, while basic categories cover the largest scope and represent the most general type. There are dependencies between basic categories, and connections can also be established between basic categories and other subcategories. Thus, operation data, data analysis data, and sensor data become different factors describing abnormal events. The classification of sensor data is similar to that of abnormal events. The specific type of various abnormal data can be determined using an expert system incorporating developer experience. Then, multiple abnormal events and sensor data are associated according to their respective types to generate various inference rules. The process from an abnormal event to the collection of required data can be abstracted as an A->S process. Here, A consists of multiple sets of abnormal-related basic categories, subcategories, and attribute data, and S consists of the sensor data to be collected. For example:

[0036] A = {<Vehicle-side software 1 (basic category), control (subcategory), steering wheel angle suddenly increases (attribute)>, <Vehicle, status, speed decreases to zero>, <Hardware, no position data on CAN bus>}

[0037] The deduced result is:

[0038] S = {<Collision caused by loss of location data, requiring the collection of video data and GPS positioning data for 10 minutes before and after the collision>}.

[0039] Example 2 of reasoning rules:

[0040] A = {<Operation, Working Condition, Transportation>, <Operation, Behavior, Emergency Braking>};

[0041] S = {<Sudden stop due to inability to identify target, data from cameras and lidar on this section of the road needs to be collected>}.

[0042] Through the above process, abnormal events and sensor data are categorized from at least four aspects: environment, operation, hardware, and software, covering multiple abnormal scenarios. Based on the direct data at the time of the anomaly, inference rules are used to deduce the cause of the anomaly and establish a mapping between the cause of the anomaly and the data that needs to be collected, thereby realizing the function of automatically collecting data on demand according to the cause of the anomaly. For example, when the vehicle's force sensor detects a sudden change in the vehicle's force signal, it is considered that a collision has occurred, and the resulting abnormal event is a collision event. According to the inference rules, the result is that in addition to collecting data from the force sensor, the system also needs to collect data from the camera, ultrasonic radar, and software commands. However, the inference rules believe that a collision event does not require the collection of navigation position data and hardware bus data, thus selectively acquiring the corresponding sensor data, forming a data packet with stronger scene adaptability, and sending the data packet to the cloud platform 2 to retrain the machine learning model.

[0043] Furthermore, in this embodiment, abnormal events are generated by manually defining the simulation platform and test data, and these defined abnormal events are stored in a historical abnormal event database. When an abnormality occurs, the inference engine first searches the historical abnormal event database for a matching abnormal event type. If a match is found, the event type is successfully identified, such as collision or sudden stop. Then, the required data is inferred according to the aforementioned inference rules, and the corresponding optimal scene data is collected by calling data sensors. If no match is found, the current event is considered a completely new abnormal event. In this case, all relevant data is collected, and after manual analysis and definition, the new abnormal event is added to the historical abnormal event database, and new inference rules are added to the rule base. This ensures the completeness of the abnormal event database and rule base, further improving the reliability of the model's abnormal event identification and rule inference.

[0044] Furthermore, in this embodiment, if the detected abnormal event is a combined event, such as simultaneously detecting a collision and an overspeed anomaly, it may be inferred that the same sensor data needs to be collected repeatedly due to the similarity of the anomalies. In order to ensure the completeness and conciseness of the optimal scene data, the sensor data is integrated before the data is uploaded to the cloud platform 2, including removing invalid data and merging the same data. Then, the integrated optimal scene data is uploaded to the cloud platform 2 for retraining, so as to further reduce the computational workload of the cloud platform 2.

[0045] Through the aforementioned components, the technical solution provided in this application establishes inference rules for abnormal data collection in response to various types of sudden anomalies in autonomous driving. When the vehicle-side machine learning model needs updating, the vehicle inputs the detected abnormal event into the rule base, matches it against multiple preset rules, and then performs abnormal data inference based on the matched inference rules. The optimal scenario data with the highest matching degree to the abnormal event is inferred from the data collected by the vehicle. This optimal scenario data is then sent to the cloud platform, allowing the cloud platform to retrain a new machine learning model for responding to abnormal events using the optimal scenario data. Finally, the vehicle receives the new machine learning model from the cloud platform and uses it to overwrite the old machine learning model stored on the vehicle for responding to abnormal events. This embodiment, based on abnormal events and inference rules, can adopt appropriate data collection strategies according to different abnormal scenarios, mine effective data for specific scenarios, and use the retrained machine learning model with effective data to make more accurate and reliable identification and decision-making regarding abnormal events. This significantly improves the reliability and effectiveness of model updates. Replacing the old machine learning model on the vehicle with the retrained machine learning model can reduce vehicle safety problems caused by decision-making errors on the vehicle.

[0046] Specifically, such as Figure 4 As shown, in one embodiment, this invention also provides a cloud-edge collaborative model update method for autonomous driving, wherein steps S101 to S103 are applied to the vehicle side, and steps S201 to S202 are applied to the cloud platform. The specific steps are as follows:

[0047] Step S101: Extract the corresponding inference rules from the rule base based on the currently detected abnormal events, and infer the optimal scenario data with the highest matching degree with the abnormal events from the data collected from the vehicle based on the inference rules.

[0048] Step S102: Send the optimal scenario data to the cloud platform.

[0049] Step S201: Receive the optimal scenario data sent by the vehicle terminal.

[0050] Step S202: Retrain a new machine learning model for dealing with abnormal events using the optimal scenario data, and send the new machine learning model to the vehicle.

[0051] Step S103: Receive the new machine learning model from the cloud platform and use the new machine learning model to overwrite the old machine learning model stored on the vehicle side for dealing with abnormal events.

[0052] Specifically, for detailed explanations of steps S101 to S103 and step S201 value S202 in this embodiment of the invention, please refer to the relevant descriptions of the above system embodiments, which will not be repeated here.

[0053] Specifically, in one embodiment, the specific steps for the vehicle-side to create inference rules in the rule base include:

[0054] Step 1: Classify multiple abnormal events into at least three categories: environment, operation, hardware, and software. Also, classify multiple sensor data into at least three categories: environment, operation, hardware, and software.

[0055] Step 2: Associate various abnormal events and sensor data according to their respective types to generate various inference rules.

[0056] In this embodiment, based on steps one and two above, step S101 specifically includes the following steps:

[0057] Step 3: Identify the event type of the abnormal event and extract the target inference rule of the same type from the rule base according to the event type.

[0058] Step 4: Acquire sensor data associated with the target inference rule and integrate the sensor data to obtain the optimal scene data.

[0059] Specifically, for a detailed explanation of steps one through four above, please refer to the relevant descriptions of the above system embodiments, which will not be repeated here.

[0060] Specifically, such as Figure 5 and Figure 6 As shown, in one embodiment, before step S101, the vehicle terminal further executes steps S104 to S107, and the cloud platform further executes steps S203 to S206. The specific steps are as follows:

[0061] Step S104: Determine whether the old machine learning model needs to be updated based on the currently detected abnormal events.

[0062] Step S105: If it is necessary to update the old machine learning model, send the data collected by the vehicle to the cloud platform.

[0063] Step S203: Receive the vehicle-mounted data collected by the vehicle terminal.

[0064] Step S204: Input the data collected by the vehicle into the source machine learning model stored on the cloud platform, and determine whether the source machine learning model can make the correct decision.

[0065] Step S205: If a correct decision cannot be made, a retraining notification is sent to the vehicle.

[0066] Step S106: If a retraining notification is received from the cloud platform, then proceed to step S101.

[0067] Step S206: If the correct decision can be made, the source machine learning model is recompressed and trimmed to obtain a retried model for dealing with abnormal events, and the retried model is sent to the vehicle.

[0068] Step S107: If a re-pruned model is received from the cloud platform, the re-pruned model is used to overwrite the old machine learning model stored on the vehicle side for dealing with abnormal events.

[0069] Specifically, to ensure the operational performance of the vehicle-side hardware and avoid excessive consumption of its computing resources, this embodiment of the invention adopts a model deployment method of deploying a large model on the cloud platform and a small model on the vehicle. The compression and pruning of the large model on the cloud platform into a small model mainly includes three steps: First, the cloud platform has a model library. Based on the vehicle type and operating scenario, the cloud platform searches for the most suitable model architecture to determine the model depth, initially meeting the vehicle's latency requirements. Second, the cloud platform compresses the searched model, reducing the number of model parameters. The goal is to prune the size of each channel and block in each layer of the neural network without sacrificing accuracy. The pruning strategy can be verified using simulation data stored on the cloud platform. Third, the weights and output values ​​of each layer in the pruned small model are further quantified to achieve lower memory consumption. Through these three steps, an optimal model that meets the vehicle's hardware resource constraints is output.

[0070] Based on this, this invention establishes an effective model evaluation and learning mechanism to achieve effective model iteration. Since model retraining inevitably consumes significant computing resources, to further conserve these resources, the vehicle-mounted system needs to determine whether an update to the machine learning model is necessary before updating the model. This includes assessing whether the vehicle can make a correct decision regarding the current abnormal event, and whether the preset model update cycle has been reached. If the vehicle-mounted system cannot make a correct decision regarding the current abnormal event or the preset model update time has been reached, the old machine learning model needs to be updated. However, if the loss of location data is due to GPS hardware malfunctions, which do not involve machine learning model identification and decision-making, then model iteration is not required.

[0071] Afterwards, the vehicle sends all the data collected by each sensor (i.e., vehicle-side data) to the cloud platform for verification. The cloud platform uses the vehicle-side data to replay abnormal events and check whether its large model (the source machine learning model) can make decisions correctly to avoid anomalies. If the source machine learning model can make decisions correctly, it means that the cloud platform's model can handle abnormal events and does not need to be retrained. This indicates that the vehicle's small model was pruned abnormally, and only the source machine learning model needs to be pruned again to obtain a small model, avoiding the needless retraining step and further saving computing resources. After the cloud platform re-optimizes and prunes to obtain a new small model (the re-pruned model), it uses the saved data to perform a second verification of the re-pruned model to determine whether it can handle abnormal events and whether its decision results are consistent with the large model. If the verification passes, the re-pruned model is fed back to the vehicle, completing the update of the vehicle's old machine learning model. If the verification fails, the source machine learning model is pruned again until the verification passes.

[0072] If the cloud platform uses data collected from the vehicle to replay an abnormal event and detects that the cloud platform's source machine learning model is also unable to make decisions normally, a manual decision can be made as to whether a model iteration is needed after this event. If the conclusion is that it is needed, then step S101 is executed. In practical applications, the system usually automatically determines that retraining is needed and sends a retraining notification to the vehicle, so that the vehicle starts to execute step S101 to perform incremental learning on the current data.

[0073] Specifically, in one embodiment, step S202 above includes the following steps:

[0074] Step 5: Retrain the new source machine learning model using the optimal scene data.

[0075] Step 6: Compress and prune the new source machine learning model to obtain a new machine learning model.

[0076] Step 7: Input the optimal scenario data into the new machine learning model and verify whether the new machine learning model can make the correct decision.

[0077] Step 8: If the correct decision can be made, retain the new machine learning model; if the correct decision cannot be made, return to the step of compressing and pruning the new source machine learning model until it is verified that the new machine learning model can make the correct decision.

[0078] Specifically, in this embodiment, based on the mechanism of deploying a large model on the cloud platform and a small model on the vehicle, when the cloud platform retrains a new source machine learning model using the optimal scenario data sent by the vehicle, the cloud platform prunes and optimizes the new source machine learning model. Similarly, the pruned small model needs to be verified. Abnormal events are replayed using data stored on the cloud platform to verify whether the small model can make correct decisions regarding abnormal events. If it can make correct decisions, the small model is retained as the new machine learning model and deployed to the vehicle to overwrite the old machine learning model. If it cannot make correct decisions, the process returns to compressing and pruning the new source machine learning model until it is verified that the new machine learning model can make correct decisions. Through this repeated pruning and verification method, the accuracy and reliability of the new machine learning model are further improved, the frequency of updates and iterations of the vehicle-side machine learning model is reduced, computing resources are saved, and the ability of the vehicle-side autonomous driving system to respond to abnormal events is improved, reducing losses caused by abnormal events in autonomous driving.

[0079] Through the above steps, the technical solution provided in this application establishes inference rules for abnormal data collection in response to various types of sudden anomalies in autonomous driving. When the machine learning model on the vehicle needs to be updated, the vehicle inputs the detected abnormal event into the rule base and matches it among various preset rules. Based on the matched inference rules, abnormal data inference is performed to deduce the optimal scenario data with the highest matching degree from the data collected on the vehicle. Then, the optimal scenario data is sent to the cloud platform, allowing the cloud platform to retrain a new machine learning model for responding to abnormal events using the optimal scenario data. Finally, the vehicle receives the new machine learning model from the cloud platform and uses it to overwrite the old machine learning model stored on the vehicle for responding to abnormal events. This embodiment, based on abnormal events and inference rules, can adopt appropriate data collection strategies according to different abnormal scenarios, mine effective data for specific scenarios, and use the retrained machine learning model with effective data to make more accurate and reliable identification and decisions regarding abnormal events. This significantly improves the reliability and effectiveness of model updates. Replacing the old machine learning model on the vehicle with the retrained machine learning model can reduce vehicle safety problems caused by decision-making errors on the vehicle.

[0080] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A cloud-edge collaborative model update method for autonomous driving, applied to the vehicle side, characterized in that, The method includes: Multiple abnormal events should be categorized at least by environment, operation, hardware, and software, and multiple sensor data should also be categorized at least by environment, operation, hardware, and software. Multiple abnormal events and multiple sensor data should be associated according to their respective categories to generate multiple inference rules and store them in the rule base. Identify the event type of the abnormal event, and extract the target inference rule of the same type from the rule base according to the event type; obtain sensor data associated with the target inference rule, and integrate the sensor data to obtain the optimal scene data with the highest matching degree with the abnormal event; Send the optimal scenario data to the cloud platform; The system receives a new machine learning model from the cloud platform and uses the new machine learning model to overwrite the old machine learning model stored on the vehicle side for dealing with the abnormal event. The new machine learning model is a machine learning model retrained by the cloud platform using the optimal scenario data for dealing with the abnormal event.

2. The method according to claim 1, characterized in that, Before extracting the corresponding inference rules from the rule base based on the currently detected abnormal events, the method further includes: Determine whether the old machine learning model needs to be updated based on the currently detected abnormal events; If the old machine learning model needs to be updated, the data collected at the vehicle end will be sent to the cloud platform. If a retraining notification is received from the cloud platform, the step of extracting the corresponding inference rules from the rule base based on the currently detected abnormal events is executed. The retraining notification is issued by the cloud platform when the source machine learning model cannot make a correct decision after the received vehicle-side collected data is input into the source machine learning model stored in the cloud platform.

3. The method according to claim 2, characterized in that, The method further includes: If a re-pruned model is received from the cloud platform, the re-pruned model is used to overwrite the old machine learning model stored on the vehicle for dealing with the abnormal event. The re-pruned model is obtained by re-compressing and pruning the source machine learning model stored on the cloud platform after the cloud platform inputs the received data collected from the vehicle into the source machine learning model stored on the cloud platform. When the source machine learning model can make the correct decision, the source machine learning model is re-compressed and pruned to obtain the machine learning model for dealing with the abnormal event.

4. A cloud-edge collaborative model update method for autonomous driving, applied to a cloud platform, characterized in that... The method includes: The optimal scene data sent by the vehicle terminal is received. The optimal scene data inference method is as follows: Multiple abnormal events are classified into types at least according to environment, operation, hardware, and software; multiple sensor data are also classified into types at least according to environment, operation, hardware, and software; multiple abnormal events and multiple sensor data are associated according to their respective types to generate multiple inference rules and store them in a rule base; the event type of the abnormal event is identified, and a target inference rule of the same type is extracted from the rule base based on the event type; sensor data associated with the target inference rule is obtained, and the sensor data is integrated to obtain the optimal scene data with the highest matching degree to the abnormal event; A new machine learning model for responding to the abnormal event is retrained using the optimal scenario data, and the new machine learning model is sent to the vehicle terminal so that the vehicle terminal can use the new machine learning model to overwrite the old machine learning model stored on the vehicle terminal.

5. The method according to claim 4, characterized in that, Before receiving the optimal scene data from the receiving vehicle, the method further includes: Receive vehicle-side collected data sent by the vehicle terminal, wherein the vehicle terminal collected data is the vehicle terminal collected data sent by the vehicle terminal when it determines that the old machine learning model needs to be updated based on the currently detected abnormal event; The vehicle-mounted data is input into the source machine learning model stored on the cloud platform, and it is determined whether the source machine learning model can make the correct decision. If the correct decision cannot be made, a retraining notification is sent to the vehicle.

6. The method according to claim 5, characterized in that, The method further includes: If the correct decision can be made, the source machine learning model is recompressed and trimmed to obtain a re-trimmed model for dealing with the abnormal event. The re-trimmed model is then sent to the vehicle terminal so that when the vehicle terminal receives the re-trimmed model, it can use the re-trimmed model to overwrite the old machine learning model stored on the vehicle terminal for dealing with the abnormal event.

7. The method according to claim 4, characterized in that, The step of retraining a new machine learning model to respond to the abnormal event using the optimal scenario data includes: The new source machine learning model was retrained using the optimal scene data; The novel machine learning model is compressed and trimmed to obtain the new machine learning model. The optimal scenario data is input into the new machine learning model to verify whether the new machine learning model can make the correct decision. If the correct decision can be made, the new machine learning model is retained; if the correct decision cannot be made, the step of compressing and pruning the new source machine learning model is returned until it is verified that the new machine learning model can make the correct decision.

8. A cloud-edge collaborative model update system for autonomous driving, characterized in that, The system includes a vehicle terminal and a cloud platform, wherein: The vehicle-side infers the optimal scene data as follows: Multiple abnormal events are categorized at least by environment, operation, hardware, and software; multiple sensor data are also categorized at least by environment, operation, hardware, and software; multiple abnormal events and multiple sensor data are associated according to their respective types to generate multiple inference rules and store them in a rule base; the event type of the abnormal event is identified, and a target inference rule of the same type is extracted from the rule base based on the event type; sensor data associated with the target inference rule is obtained, and the sensor data is integrated to obtain the optimal scene data with the highest matching degree to the abnormal event; The vehicle sends the optimal scenario data to the cloud platform; The cloud platform receives the optimal scenario data sent by the vehicle and uses the optimal scenario data to retrain a new machine learning model to deal with the abnormal event. The cloud platform sends the new machine learning model to the vehicle. The vehicle receives a new machine learning model from the cloud platform and uses the new machine learning model to overwrite the old machine learning model stored on the vehicle for dealing with the abnormal event.

Images