Vehicle fault prediction method and device and vehicle maintenance system
By using a joint prediction method of public models and on-board models in vehicle fault detection, combined with time-dependent weight calculation and regular update of on-board models, the problems of low accuracy and high cost in the prior art are solved, and higher fault prediction accuracy and lower maintenance waste are achieved.
Patent Information
- Application Number
- CN202311842214.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-28
- Publication Date
- 2025-07-01
AI Technical Summary
The existing vehicle fault detection methods have lost money when detecting faults, the accuracy of predictive maintenance is not high, the recall rate is low, and the model needs frequent retraining, which is costly.
By using a joint prediction method of public models and on-board models, the accuracy of the model prediction of vehicle failures is improved. The public model is trained by training data collected within the preset time range. The weight coefficient is related to the time of the training data. The on-board model is updated regularly and the weight is adjusted according to the actual operation of the user.
It improves the accuracy of vehicle failure prediction, reduces false predictions, avoids unnecessary repairs caused by false predictions, reduces waste of manpower, material resources and time, and improves user experience.
Smart Images
Figure CN120235602A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of vehicle intelligence, and particularly to a vehicle fault prediction method, device, and vehicle maintenance system. Background Art
[0002] With the development of technology and the process of urbanization, there are more and more vehicles on the road. In recent years, vehicle intelligence has also developed rapidly, and the demand for vehicle maintenance has been increasing.
[0003] Currently, the realization of vehicle maintenance includes the following two methods:
[0004] The first method detects vehicle faults according to vehicle data through preset rules.
[0005] The second method makes predictions through an artificial intelligence (AI) model to predict future faults to a certain extent, that is, predictive maintenance.
[0006] Among them, predictive maintenance is a new direction for maintaining mechanical components, electronic components, etc. Its main function is to be able to predict the occurrence of faults and obstacles before they occur, so that corresponding measures can be taken before losses occur to reduce losses.
[0007] It should be noted that the above introduction to the technical background is only for the convenience of clearly and completely explaining the technical solution of the present invention and facilitating the understanding of those skilled in the art. It cannot be considered that the above technical solutions are well-known to those skilled in the art just because these solutions are described in the background art part of the present invention. Summary of the Invention
[0008] However, for the existing first method, when a fault is detected, it means that the fault has occurred, and thus the loss has occurred; for the existing second method, the prediction accuracy is not high, the recall is not high, and as time goes by, the model needs to be retrained, with a relatively high cost. In addition, when training the model, the weights of the training data used are the same, and past data cannot better generalize the current situation. Due to the passage of time, past data has an advantage in terms of quantity, resulting in a relatively large overall weight of past data, thus leading to a decrease in prediction accuracy.
[0009] To solve one or more of the above problems in the prior art, embodiments of the present invention provide a vehicle fault prediction method, device, and vehicle maintenance system. By improving the prediction accuracy of the model for vehicle faults, avoiding or reducing incorrect fault predictions, thus avoiding or reducing the waste of manpower, material resources, and time caused by users performing vehicle maintenance on the vehicle repair structure due to incorrect fault predictions, and improving the user experience.
[0010] According to the first aspect of the embodiments of the present invention, a vehicle fault prediction method is provided. The method includes: training a common model to obtain a trained common model; inputting vehicle data of a target vehicle into the trained common model to output a first prediction result regarding vehicle faults; inputting the vehicle data and the first prediction result into an in-vehicle model of the target vehicle to output a second prediction result regarding vehicle faults; and giving a reminder to the user of the target vehicle according to the second prediction result. When training the common model, training is performed using training data collected within a preset time range, and a weight coefficient for calculating the weight of the training data is related to the time of the training data.
[0011] According to the second aspect of the embodiments of the present invention, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient, and the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient.
[0012] According to the third aspect of the embodiments of the present invention, when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1.
[0013] According to the fourth aspect of the embodiments of the present invention, when training the common model, the recall rate at the previous moment is used as the weight of the accuracy rate at the current moment, and the accuracy rate at the previous moment is used as the weight of the recall rate at the current moment to calculate the score at the current moment.
[0014] According to the fifth aspect of the embodiments of the present invention, the weight of the training data is calculated based on the weight coefficient and a corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient. When the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
[0015] According to the sixth aspect of the embodiments of the present invention, the method further includes: regularly updating the in-vehicle model of the target vehicle. During the update process of the in-vehicle model, the first in-vehicle model and the second in-vehicle model are kept running. The first in-vehicle model is the original model, and the second in-vehicle model is the newly issued model.
[0016] According to the seventh aspect of the embodiments of the present invention, during the update process of the vehicle model, when the output results of the first vehicle model are the same as those of the second vehicle model, the output results of the first vehicle model are used as the second prediction result; when the output results of the first vehicle model are different from those of the second vehicle model and the weight of the first vehicle model is greater than the weight of the second vehicle model, the output results of the first vehicle model are used as the second prediction result; when the output results of the first vehicle model are different from those of the second vehicle model and the weight of the first vehicle model is less than or equal to the weight of the second vehicle model, the output results of the second vehicle model are used as the second prediction result.
[0017] According to the eighth aspect of the embodiments of the present invention, during the update process of the vehicle model, the weight of the first vehicle model is updated according to whether the output result of the first vehicle model conforms to the actual operation of the user, and the weight of the second vehicle model is updated according to whether the output result of the second vehicle model conforms to the actual operation of the user.
[0018] According to the ninth aspect of the embodiments of the present invention, during the update process of the vehicle model, when a new third vehicle model is received before the update is completed, the third vehicle model is used to replace the second vehicle model for the update.
[0019] According to the tenth aspect of the embodiments of the present invention, the method further includes: adjusting the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder.
[0020] According to the eleventh aspect of the embodiments of the present invention, a vehicle fault prediction device is provided. The device includes: a training module that trains a public model to obtain a trained public model; a prediction module that inputs vehicle data of a target vehicle into the trained public model and outputs a first prediction result regarding vehicle faults; inputs the vehicle data and the first prediction result into the vehicle model of the target vehicle and outputs a second prediction result regarding vehicle faults; and a reminder module that reminds the user of the target vehicle according to the second prediction result. When the training module trains the public model, it uses training data collected within a preset time range for training, and the weight coefficient for calculating the weight of the training data is related to the time of the training data.
[0021] According to the twelfth aspect of the embodiments of the present invention, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient; the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient. And when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1.
[0022] According to the thirteenth aspect of the embodiments of the present invention, when the training module trains the common model, it uses the recall rate at the previous moment as the weight of the accuracy rate at the current moment, and uses the accuracy rate at the previous moment as the weight of the recall rate at the current moment to calculate the score at the current moment.
[0023] According to the fourteenth aspect of the embodiments of the present invention, the weight of the training data is calculated based on the weight coefficient and the corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient; when the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
[0024] According to the fifteenth aspect of the embodiments of the present invention, the device further includes: an update module that regularly updates the in-vehicle model of the target vehicle. During the update process of the in-vehicle model, the first in-vehicle model and the second in-vehicle model are kept running. The first in-vehicle model is the original model, and the second in-vehicle model is the newly issued model.
[0025] According to the sixteenth aspect of the embodiments of the present invention, during the update process of the in-vehicle model, when the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, the output result of the second in-vehicle model is used as the second prediction result.
[0026] According to the seventeenth aspect of the embodiments of the present invention, during the update process of the in-vehicle model, the update module updates the weight of the first in-vehicle model according to whether the output result of the first in-vehicle model conforms to the actual operation of the user, and updates the weight of the second in-vehicle model according to whether the output result of the second in-vehicle model conforms to the actual operation of the user.
[0027] According to the eighteenth aspect of the embodiments of the present invention, during the update process of the in-vehicle model, when a new third in-vehicle model is received before the update is completed, the third in-vehicle model is used to replace the second in-vehicle model for the update.
[0028] According to the nineteenth aspect of the embodiments of the present invention, the reminder module adjusts the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder.
[0029] According to the twentieth aspect of the embodiments of the present invention, a vehicle maintenance system is provided. The system includes: a maintenance structure for maintaining the vehicle; a server storing a trained public model, which is trained using training data collected from the maintenance structure within a preset time range during the training of the public model, and the weight coefficient for calculating the weight of the training data is related to the time of the training data; an in-vehicle system storing an in-vehicle model; the server inputs the vehicle data of the target vehicle into the trained public model to output a first prediction result regarding vehicle faults; and sends the first prediction result to the in-vehicle system of the target vehicle. The in-vehicle system of the target vehicle inputs the vehicle data and the first prediction result into the in-vehicle model to output a second prediction result regarding vehicle faults, and reminds the user of the target vehicle according to the second prediction result.
[0030] One of the beneficial effects of the embodiments of the present invention is as follows:
[0031] On the basis of preliminarily predicting vehicle faults using a public model based on big data, the in-vehicle model of the vehicle is further used to perform refined prediction of the faults of this vehicle, thereby improving the accuracy of fault prediction.
[0032] Moreover, during the training of the public model, the weight of the training data is associated with the time of the training data, so that the training data can more reasonably reflect the current situation, thereby improving the prediction accuracy of the trained public model.
[0033] Therefore, by improving the prediction accuracy of the model for vehicle faults, avoiding or reducing errors in fault prediction, the waste of manpower, material resources, and time caused by the user going to the maintenance structure for vehicle maintenance due to incorrect fault prediction is avoided or reduced, and the user experience is improved.
[0034] Furthermore, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient; the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient. And when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1. Thus, reasonable weights can be assigned to each training data according to time, further improving the prediction accuracy of the common model.
[0035] Furthermore, when training the common model, the recall rate at the previous moment is used as the weight of the accuracy rate at the current moment, and the accuracy rate at the previous moment is used as the weight of the recall rate at the current moment to calculate the score at the current moment. In this way, the accuracy rate (accuracy) and recall rate (recall) during model training can be reasonably balanced, and past characteristics can be brought into the new iteration. Further improving the prediction accuracy and training efficiency of the common model.
[0036] Furthermore, the weight of the training data is calculated based on the weight coefficient and the corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient; when the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero. In this way, a corruption mechanism is introduced when calculating the weight of the training data, enabling full consideration of negative feedback (bad case), and further improving the prediction accuracy of the common model.
[0037] Furthermore, the in-vehicle model of the target vehicle is updated regularly. During the update process of the in-vehicle model, the original first in-vehicle model and the new second in-vehicle model are kept running. In this way, when the new model has not yet achieved the expected function, the old model will still play a role. Thus, the model can be replaced smoothly without the user's perception, solving problems such as the user's discomfort during the update of the in-vehicle model and the overly long adaptation time caused by cold start.
[0038] Furthermore, during the update process of the in-vehicle model, when the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, the output result of the second in-vehicle model is used as the second prediction result. In this way, by comparing the weights of the old and new models to determine which model's prediction result to use, the accuracy of the prediction result during the model update process can be ensured.
[0039] Further, during the update process of the vehicle model, according to whether the output result of the first vehicle model conforms to the actual operation of the user, the weight of the first vehicle model is updated, and according to whether the output result of the second vehicle model conforms to the actual operation of the user, the weight of the second vehicle model is updated. In this way, the weights of the old and new models are updated according to whether their output results conform to the actual operation of the user, that is, the effectiveness of the old and new models is verified according to the actual operation and their respective weights are adjusted, which can ensure the rationality of the model update speed and update degree during the model process.
[0040] Further, during the update process of the vehicle model, when a new third vehicle model is received before the update is completed, the third vehicle model is used to replace the second vehicle model for the update. In this way, even if a new model is issued during the model update process, the replacement of the old and new models can be successfully completed without the user's perception.
[0041] Further, according to whether the user has performed relevant operations on the reminder, the time interval and reminder probability of the next reminder are adjusted. In this way, by introducing a penalty mechanism in the reminder and adjusting the time interval and possibility of the reminder according to the user's behavior, the reminder for predicting faults can better conform to the user's preferences, avoid disturbing the user, and further improve the user experience.
[0042] With reference to the following description and drawings, the embodiments of the present invention are disclosed in detail. It should be understood that the embodiments of the present invention are not limited in scope thereby. Within the spirit and terms of the appended claims, the embodiments of the present invention include many changes, modifications and equivalents.
[0043] Features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, combined with features in other embodiments, or substituted for features in other embodiments.
[0044] It should be emphasized that the term "comprising / including / having" as used herein refers to the presence of features, wholes, or components, but does not exclude the presence or addition of one or more other features, wholes or components. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] From the following detailed description in conjunction with the drawings, the above and other objects, features and advantages of the embodiments of the present invention will become more obvious. In the drawings:
[0046] Figure 1 is a schematic diagram of the vehicle fault prediction method according to Embodiment 1 of the present invention;
[0047] Figure 2 It is a configuration diagram of the vehicle maintenance system according to Embodiment 1 of the present invention;
[0048] Figure 3 It is a flowchart of the update process of the in-vehicle model according to Embodiment 1 of the present invention;
[0049] Figure 4 It is a schematic diagram of the vehicle fault prediction device according to Embodiment 2 of the present invention. Detailed implementation manners
[0050] Referring to the accompanying drawings, through the following description, the foregoing and other features of the present invention will become apparent. In the description and drawings, specific embodiments of the present invention are disclosed, which show some embodiments in which the principles of the present invention can be adopted. It should be understood that the present invention is not limited to the described embodiments, but includes all modifications, variations, and equivalents falling within the scope of the appended claims.
[0051] The vehicle fault prediction method according to the embodiments of the present invention will be described below with reference to the accompanying drawings.
[0052] Embodiment 1
[0053] Embodiment 1 of the present invention provides a vehicle fault prediction method.
[0054] Figure 1 It is a schematic diagram of the vehicle fault prediction method according to Embodiment 1 of the present invention. As Figure 1 shown, the method includes:
[0055] Step 101: Train a common model to obtain a trained common model;
[0056] Among them, when training the common model, training is performed using the training data collected within a preset time range, and the weight coefficient used to calculate the weight of the training data is related to the time of the training data.
[0057] Step 102: Input the vehicle data of the target vehicle into the trained common model, and output a first prediction result regarding vehicle faults;
[0058] Step 103: Input the vehicle data and the first prediction result into the in-vehicle model of the target vehicle, and output a second prediction result regarding vehicle faults; and
[0059] Step 104: Remind the user of the target vehicle according to the second prediction result.
[0060] In some embodiments, the method can be applied to the fault prediction of various vehicles.
[0061] The vehicle can be various types of vehicles, for example, a car, an electric vehicle, or a hybrid vehicle; for another example, a private car, a bus, a coach, a truck, a lorry, a motorcycle, a tour bus, etc.; for yet another example, a driverless car or an autonomous driving car, etc.
[0062] In some embodiments, the method can be used for predicting various faults of a vehicle.
[0063] For example, predicting various faults of each mechanical component, electronic component, and circuit connection of the vehicle. For example, predicting the battery fault of the vehicle.
[0064] In some embodiments, the target vehicle refers to a vehicle to which the vehicle fault prediction method of the embodiments of the present invention is applied.
[0065] In some embodiments, the method can be applied to a vehicle maintenance system. The vehicle fault prediction method of the embodiments of the present invention will be specifically described below in conjunction with the vehicle maintenance system.
[0066] Figure 2 is a block diagram of a vehicle maintenance system according to Embodiment 1 of the present invention. As Figure 2 shown, the vehicle maintenance system 10 includes: a maintenance structure 11 that repairs the vehicle; a server 12 that stores a trained common model, and when training the common model, uses training data collected from the maintenance structure within a preset time range for training, and the weight coefficient for calculating the weight of the training data is related to the time of the training data; a vehicle-mounted system 131 that stores a vehicle-mounted model;
[0067] The server 12 inputs the vehicle data of the target vehicle 13 into the trained common model, outputs a first prediction result regarding vehicle faults; and sends the first prediction result to the vehicle-mounted system 131 of the target vehicle 13. The vehicle-mounted system 131 of the target vehicle 13 inputs the vehicle data and the first prediction result into the vehicle-mounted model, outputs a second prediction result regarding vehicle faults, and gives a reminder to the user of the target vehicle according to the second prediction result.
[0068] For example, the server 12 is a cloud server that collects a large amount of maintenance record data of other vehicles from a maintenance institution 12 as training data.
[0069] For example, the maintenance institution 12 is various types of maintenance institutions, for example, a vehicle 4S store, a vehicle repair shop, or other types of vehicle maintenance institutions.
[0070] In this way, on the basis of initially predicting vehicle faults using a big data-based common model, the vehicle-mounted model of the vehicle is further used to perform refined prediction of the faults of this vehicle, thereby being able to improve the accuracy of fault prediction;
[0071] Moreover, in the training of the common model, the weight of the training data is associated with the time of the training data, so that the training data can more reasonably reflect the current situation, thereby improving the prediction accuracy of the trained common model.
[0072] Therefore, by improving the prediction accuracy of the model for vehicle faults, the errors in fault prediction are avoided or reduced, thereby avoiding or reducing the waste of manpower, material resources and time caused by users performing vehicle maintenance on the repair structure due to incorrect fault prediction, and improving the user experience.
[0073] In step 101, the common model is trained to obtain a trained common model.
[0074] When training the common model, the training data collected within a preset time range is used for training, and the weight coefficient used to calculate the weight of the training data is related to the time of the training data.
[0075] In some embodiments, the server 12 collects the repair record data from the maintenance agency 12 as training data. For example, the server 12 can collect the repair record data of a large number of other vehicles from each maintenance agency 12 as training data, and calibrate the time of the repair record data.
[0076] For example, the repair record data includes one or more of time, vehicle brand, vehicle model, vehicle operation data, vehicle status data, vehicle environment data, fault reminder data, fault type data, repair operation data, etc.
[0077] In some embodiments, the preset time range for collecting the training data can be set according to actual needs. For example, the repair record data within the last three months from the current time is used as the training data.
[0078] In some embodiments, the weight coefficient used to calculate the weight of the training data is related to the time of the training data. For example, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient, and the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient.
[0079] In addition, when the time interval between the time of the training data and the current time is less than the preset value, the weight coefficient is 1.
[0080] For example, training data that is beyond a preset time from the current has little impact on the current situation. The greater the time distance, the smaller the effect. For example, training data within 1 month from the current has a great impact on the current situation, data beyond 3 months from the current is useless data, and for data within 1 - 3 months from the current, the greater the time distance, the smaller the impact.
[0081] For example, the weight coefficient can be calculated by the following formula (1):
[0082]
[0083] Wherein, represents the weight coefficient of the i-th training data at time t j t represents the time, t j represents the past time, t represents the current time, and a and b are preset values.
[0084] It can be seen from formula (1) that when the time interval |t j - t| between the time of the training data and the current time is less than the preset value a, the training data has a great impact on the current situation, and its weight coefficient is 1; when the time interval |t j - t| is between the preset values a and b, it means that the training data has a certain impact on the current situation, and the greater this time interval, the smaller the impact effect, and its weight varies between 0 and 1; when the time interval |t j - t| is greater than the preset value b, the training data has little impact on the current situation, and its weight is 0.
[0085] Thus, reasonable weights can be assigned to each training data according to time, further improving the prediction accuracy of the common model.
[0086] In some embodiments, when training the common model, the recall rate at the previous moment is used as the weight of the accuracy rate at the current moment, and the accuracy rate at the previous moment is used as the weight of the recall rate at the current moment to calculate the score at the current moment. In this way, the accuracy (accuracy) and recall rate (recall) during model training can be reasonably balanced, and the past characteristics can be brought into the new iteration. The prediction accuracy and training efficiency of the common model are further improved.
[0087] For example, a weight model is used for gaming, and the previous situation can be used as the optimization basis for the next one. The calculation formula for this score is as follows:
[0088] Score t =W recall,t-1 *accuracy t +W accuracy,t-1 *recallt (2)
[0089] Among them, Score t represents the score at the current time, and W recall,t-1 represents the recall rate at the previous time and serves as the weight of the accuracy at the current time t of accuracy accuracy,t-1 represents the accuracy at the previous time and serves as the weight of the recall rate at the current time t of recall
[0090] When training the public model, the goal of optimizing the model is to increase the score Score, and the iteration ends when the score Score decreases. In this way, the recall rate and accuracy are balanced, and past characteristics are brought into the new iteration.
[0091] In some embodiments, a corruption mechanism based on bad cases is introduced when calculating the weights of the training data. For example, the calculation formula for the corruption coefficient is as follows:
[0092]
[0093] Among them, D i,t represents the corruption coefficient of the i-th training data at the current time, and v i represents the i-th training data, and v j represents the j-th bad case. Since there are many bad cases, the corruption coefficient of each training data is the sum of its performances in all bad cases.
[0094] In some embodiments, the weights of the training data are calculated based on the weight coefficient and the corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient. When the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
[0095] For example, the weights of the training data are calculated according to the following formula:
[0096]
[0097] Among them, F i,t represents the weight of the i-th training data at the current time, represents t j the weight coefficient of the i-th training data at time, and D i,t represents the corruption coefficient of the i-th training data at the current time. For and D i,t the calculation methods can refer to the above formulas (1) and (3) for example.
[0098] In this way, a corruption mechanism is introduced when calculating the weights of the training data, enabling full consideration of negative feedback (bad cases) and further improving the prediction accuracy of the public model.
[0099] In addition, when training the public model, the goal of optimizing the model is to increase the score Score, and the iteration ends when the score Score decreases. For example, the calculation method of the score Score can refer to the above formula (2).
[0100] In some embodiments, the training process of the public model can also refer to related technologies, which will not be elaborated here.
[0101] In some embodiments, during the use of the public model, the public model can also be updated. The specific update method can refer to the above training process, which will not be elaborated here.
[0102] After obtaining the trained public model, when it is necessary to predict the faults of the target vehicle, in steps 102 - 104, the vehicle data of the target vehicle is input into the trained public model to output the first prediction result regarding vehicle faults; the vehicle data and the first prediction result are input into the in-vehicle model of the target vehicle to output the second prediction result regarding vehicle faults; and according to the second prediction result, a reminder is sent to the user of the target vehicle.
[0103] In some embodiments, the first prediction result output by the public model is a prediction result based on big data, reflecting the preferences of the public.
[0104] This first prediction result is input into the in-vehicle model of the target vehicle as a preliminary prediction result for further prediction. For example, the first prediction result is a binary prediction result, that is, it indicates whether there is a fault. For example, 0 indicates no fault and 1 indicates a fault.
[0105] In the in-vehicle model, based on the input first prediction result from the public model and the vehicle data, a more accurate second prediction result is output, and this second prediction result is represented by, for example, the probability of a fault occurring.
[0106] When the probability of a fault occurring in the second prediction result exceeds a preset value, for example, exceeds 0.5, a reminder is sent to the user through the vehicle-mounted system.
[0107] For example, relevant fault reminder information is displayed on the in-vehicle terminal, such as "The battery may be faulty. Please go to a repair agency for inspection". In addition, relevant reminder information can also be pushed to the user's mobile terminal. The embodiments of the present invention do not limit the specific reminder method.
[0108] In some embodiments, the vehicle fault prediction method of the embodiments of the present invention may further include:
[0109] Regularly update the in-vehicle model of the target vehicle. For example, set the update period of the in-vehicle model according to actual needs;
[0110] During the update process of the in-vehicle model, keep the first in-vehicle model and the second in-vehicle model running. The first in-vehicle model is the original model, and the second in-vehicle model is the newly issued model, such as the new model issued by server 12.
[0111] In this way, when the new model has not yet achieved the expected function, the old model will still play a role. Thus, the model can be smoothly replaced without the user's perception, solving problems such as the user's discomfort during in-vehicle model update and the overly long adaptation time caused by cold start.
[0112] In some embodiments, during the update process of the in-vehicle model,
[0113] When the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, use the output result of the first in-vehicle model as the second prediction result;
[0114] When the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, use the output result of the first in-vehicle model as the second prediction result;
[0115] When the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, use the output result of the second in-vehicle model as the second prediction result.
[0116] For example, determine which model's output result to use according to the following formula:
[0117]
[0118] Wherein, Result represents the second prediction result output by the in-vehicle model, result old represents the output result of the original first in-vehicle model, result new represents the output result of the newly issued second in-vehicle model, W old represents the weight of the original first in-vehicle model, W new represents the weight of the newly issued second in-vehicle model.
[0119] In addition, during the update process of the in-vehicle model, the weight W of the original first in-vehicle model oldThe initial value of is 1, and the weight W of the newly issued second vehicle-mounted model new has an initial value of 0.
[0120] In this way, by comparing the weights of the old and new models to determine which model's prediction result to use, the accuracy of the prediction result during the model update process can be ensured.
[0121] In some embodiments, during the update process of the vehicle-mounted model, according to whether the output result of the first vehicle-mounted model conforms to the actual operation of the user, the weight of the first vehicle-mounted model is updated, and according to whether the output result of the second vehicle-mounted model conforms to the actual operation of the user, the weight of the second vehicle-mounted model is updated.
[0122] For example, the weights of the original first vehicle-mounted model and the newly issued second vehicle-mounted model are updated according to the following formula:
[0123]
[0124] where W old,t is the weight of the original first vehicle-mounted model at the current time, W new,t is the weight of the newly issued second vehicle-mounted model at the current time, result old,t represents the output result of the original first vehicle-mounted model at the current time, result new,t represents the output result of the newly issued second vehicle-mounted model at the current time, operation t represents the actual operation at the current time, that is, for the occurrence of a fault, whether the user has performed relevant operations, such as going to a maintenance agency for repair or performing fault troubleshooting by oneself to eliminate the fault.
[0125] In this way, by updating the respective weights according to whether the output results of the old and new models conform to the actual operation of the user, that is, verifying the accuracy and effectiveness of the old and new models according to the actual operation and adjusting their respective weights, the rationality of the model update speed and update degree during the model process can be ensured.
[0126] In some embodiments, during the update process of the vehicle-mounted model, when a newly issued third vehicle-mounted model is received before the update is completed, the third vehicle-mounted model is used to replace the second vehicle-mounted model for the update.
[0127] For example, before the weight of the newly issued second vehicle-mounted model is still less than the weight of the original first vehicle-mounted model, a new vehicle-mounted model is issued from the server.
[0128] Figure 3 is a flowchart of the update process of the vehicle-mounted model in Embodiment 1 of the present invention. As Figure 3As shown in the figure, the update process of the vehicle-mounted model includes:
[0129] Step 301: Send down a new second vehicle-mounted model;
[0130] Step 302: Set the weight initial value of the original first vehicle-mounted model to 1, and set the weight initial value of the newly sent-down second vehicle-mounted model to 0;
[0131] Step 303: Calculate the output result according to the weights of the first vehicle-mounted model and the second vehicle-mounted model;
[0132] Step 304: Determine whether the output results of the first vehicle-mounted model and the second vehicle-mounted model conform to the user's actual operation;
[0133] Step 305: Update the weights of the first vehicle-mounted model and the second vehicle-mounted model according to the judgment result;
[0134] Step 306: Determine whether the weight of the first vehicle-mounted model is less than the weight of the second vehicle-mounted model; when the judgment result is "yes", go to Step 307, and when the judgment result is "no", go to Step 308;
[0135] Step 307: Stop using the first vehicle-mounted model and use the second vehicle-mounted model;
[0136] Step 308: Determine whether there is a newly sent-down third vehicle-mounted model; when the judgment result is "yes", go to Step 309, and when the judgment result is "no", go to Step 303;
[0137] Step 309: Use the third vehicle-mounted model as the new second vehicle-mounted model to continue the update, and stop using the previous second vehicle-mounted model.
[0138] In this way, even if a new model is sent down during the model update process, the replacement of the old and new models can be successfully completed smoothly without the user's perception.
[0139] In some embodiments, the method further includes: adjusting the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder. For example, the relevant operations include operations such as going to a maintenance agency for repair or performing a fault diagnosis by oneself to eliminate the fault.
[0140] For example, a penalty mechanism is introduced in the reminder. For example, the time interval of the reminder is adjusted through the following formula:
[0141]
[0142] where t represents the time interval between the next reminder and the current reminder, t intervalIt represents the set reminder time interval. repair = 0 indicates that the user has not performed relevant operations after being reminded, and n represents the number of times the user has not performed relevant operations after being reminded.
[0143] For example, adjust the reminder probability through the following formula:
[0144]
[0145] Among them, P i represents the current reminder probability, P i-1 represents the probability of the previous reminder. repair = 0 indicates that the user has not performed relevant operations after being reminded, and n represents the number of times the user has not performed relevant operations after being reminded. For example, a reminder is made when the reminder probability is greater than 0.5.
[0146] It can be seen that if the user does not perform relevant operations such as repair after being reminded, then the reminder time interval will be increased and the reminder probability will be decreased.
[0147] In this way, by introducing a penalty mechanism in the reminder and adjusting the reminder time interval and possibility according to the user's behavior, the reminder for predicted faults can better meet the preferences of this user, avoid disturbing the user, and further improve the user experience.
[0148] As can be seen from the above embodiments, on the basis of initially predicting vehicle faults using a public model based on big data, further use the vehicle's in-vehicle model to perform refined prediction of the faults of this vehicle, thereby improving the accuracy of fault prediction;
[0149] Moreover, in the training of the public model, an association is established between the weight of the training data and the time of the training data, so that the training data can more reasonably reflect the current situation, thereby improving the prediction accuracy of the trained public model;
[0150] Therefore, by improving the prediction accuracy of the model for vehicle faults, avoiding or reducing incorrect fault predictions, thereby avoiding or reducing the waste of manpower, material resources, and time caused by the user's vehicle repair due to incorrect fault predictions, and improving the user experience.
[0151] Furthermore, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient; the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient. And when the time interval between the time of the training data and the current time is less than the preset value, the weight coefficient is 1. Thus, reasonable weights can be assigned to each training data according to time, further improving the prediction accuracy of the public model.
[0152] Further, when training the common model, the recall rate at the previous moment is used as the weight of the accuracy rate at the current moment, and the accuracy rate at the previous moment is used as the weight of the recall rate at the current moment to calculate the score at the current moment. In this way, it is possible to reasonably balance the accuracy rate and recall rate during model training and bring past characteristics into new iterations, further improving the prediction accuracy and training efficiency of the common model.
[0153] Further, the weight of the training data is calculated based on the weight coefficient and the corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient. When the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero. In this way, a corruption mechanism is introduced when calculating the weight of the training data, enabling full consideration of negative feedback (bad cases) and further improving the prediction accuracy of the common model.
[0154] Further, the in-vehicle model of the target vehicle is updated regularly. During the update process of the in-vehicle model, the original first in-vehicle model and the new second in-vehicle model are kept running. In this way, when the new model has not yet achieved the expected function, the old model will still play a role, so that the model can be smoothly replaced without the user's perception, solving problems such as user discomfort and overly long adaptation time caused by cold start during in-vehicle model update.
[0155] Further, during the update process of the in-vehicle model, when the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, the output result of the second in-vehicle model is used as the second prediction result. In this way, by comparing the weights of the old and new models to determine which model's prediction result to use, the accuracy of the prediction result during the model update process can be guaranteed.
[0156] Further, during the update process of the vehicle model, the weights of the first vehicle model are updated according to whether the output result of the first vehicle model conforms to the actual operation of the user, and the weights of the second vehicle model are updated according to whether the output result of the second vehicle model conforms to the actual operation of the user. In this way, the weights of the old and new models are updated according to whether their output results conform to the actual operation of the user, that is, the effectiveness of the old and new models is verified according to the actual operation and their respective weights are adjusted, which can ensure the rationality of the model update speed and update degree during the model process.
[0157] Further, during the update process of the vehicle model, when a new third vehicle model is received before the update is completed, the third vehicle model is used to replace the second vehicle model for the update. In this way, even if a new model is issued during the model update process, the replacement of the old and new models can be smoothly completed without the user's perception.
[0158] Further, according to whether the user has performed relevant operations on the reminder, the time interval and reminder probability of the next reminder are adjusted. In this way, by introducing a penalty mechanism in the reminder and adjusting the time interval and possibility of the reminder according to the user's behavior, the reminder of the predicted fault can better conform to the user's preferences, avoid disturbing the user, and further improve the user experience.
[0159] Embodiment 2
[0160] Embodiment 2 of the present invention provides a vehicle fault prediction device. This system corresponds to the vehicle fault prediction method described in Embodiment 1. For specific content, reference can be made to the description in Embodiment 1, and repeated content will not be specifically described again.
[0161] Figure 4 is a schematic diagram of the vehicle fault prediction device according to Embodiment 2 of the present invention. As Figure 4 shown, the vehicle fault prediction device 400 includes:
[0162] A training module 401 that trains a common model to obtain a trained common model;
[0163] A prediction module 402 that inputs the vehicle data of the target vehicle into the trained common model and outputs a first prediction result regarding vehicle faults; inputs the vehicle data and the first prediction result into the vehicle model of the target vehicle and outputs a second prediction result regarding vehicle faults; and
[0164] A reminder module 403 that reminds the user of the target vehicle according to the second prediction result.
[0165] Among them, when training the common model, the training module 401 uses the training data collected within a preset time range for training, and the weight coefficient used to calculate the weight of the training data is related to the time of the training data.
[0166] In some embodiments, the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient; the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient.
[0167] In some embodiments, when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1.
[0168] In some embodiments, when training the common model, the training module 401 uses the recall rate at the previous moment as the weight of the accuracy rate at the current moment, and uses the accuracy rate at the previous moment as the weight of the recall rate at the current moment to calculate the score at the current moment.
[0169] In some embodiments, the weight of the training data is calculated based on the weight coefficient and the corruption coefficient. When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient; when the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
[0170] In some embodiments, as Figure 4 shown, the device 400 further includes:
[0171] An update module 404 that periodically updates the in-vehicle model of the target vehicle. During the update process of the in-vehicle model, the operation of the first in-vehicle model and the second in-vehicle model is maintained. The first in-vehicle model is the original model, and the second in-vehicle model is the newly issued model.
[0172] In some embodiments, during the update process of the in-vehicle model, when the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, the output result of the second in-vehicle model is used as the second prediction result.
[0173] In some embodiments, during the update process of the vehicle-mounted model, the update module 404 updates the weights of the first vehicle-mounted model according to whether the output result of the first vehicle-mounted model conforms to the actual operation of the user, and updates the weights of the second vehicle-mounted model according to whether the output result of the second vehicle-mounted model conforms to the actual operation of the user.
[0174] In some embodiments, during the update process of the vehicle-mounted model, when a new third vehicle-mounted model is received before the update is completed, the update module 404 uses the third vehicle-mounted model to replace the second vehicle-mounted model for update.
[0175] In some embodiments, the reminder module 403 adjusts the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder.
[0176] As can be seen from the above embodiments, on the basis of using a public model based on big data to preliminarily predict vehicle faults, the vehicle-mounted model of the vehicle is further used to perform refined prediction on the faults of the vehicle, so as to improve the accuracy of fault prediction;
[0177] Moreover, in the training of the public model, the weights of the training data are associated with the time of the training data, so that the training data can more reasonably reflect the current situation, thereby improving the prediction accuracy of the trained public model;
[0178] Therefore, by improving the prediction accuracy of the model for vehicle faults, the errors in fault prediction are avoided or reduced, thereby avoiding or reducing the waste of manpower, material resources and time caused by the user's vehicle repair due to incorrect fault prediction, and improving the user experience.
[0179] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, indicating the ways in which the principles of the present invention can be adopted. However, it should be understood that the implementation of the present invention is not limited to the above embodiments, and also includes all changes, modifications and equivalents that do not deviate from the scope of the present invention.
Claims
1. A vehicle fault prediction method, characterized in that, The method includes: Training a common model to obtain a trained common model; Inputting the vehicle data of the target vehicle into the trained common model to output a first prediction result regarding vehicle faults; Inputting the vehicle data and the first prediction result into the in-vehicle model of the target vehicle to output a second prediction result regarding vehicle faults; and Reminding the user of the target vehicle according to the second prediction result, wherein, when training the common model, training is performed using training data collected within a preset time range, and the weight coefficient used to calculate the weight of the training data is related to the time of the training data.
2. The method according to claim 1, wherein the longer the time interval between the time of the training data and the current time, the smaller the weight coefficient, the shorter the time interval between the time of the training data and the current time, the larger the weight coefficient.
3. The method according to claim 1 or 2, wherein when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1.
4. The method according to claim 1, wherein when training the common model, the recall rate at the previous moment is used as the weight of the accuracy rate at the current moment, and the accuracy rate at the previous moment is used as the weight of the recall rate at the current moment to calculate the score at the current moment.
5. The method according to claim 1, wherein the weight of the training data is calculated according to the weight coefficient and the corruption coefficient, when the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient, when the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
6. The method according to claim 1, wherein The method further includes: Regularly updating the in-vehicle model of the target vehicle, during the update process of the in-vehicle model, keeping the first in-vehicle model and the second in-vehicle model running, where the first in-vehicle model is the original model and the second in-vehicle model is the newly issued model.
7. The method according to claim 6, wherein during the update process of the in-vehicle model, when the output result of the first in-vehicle model is the same as the output result of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is greater than the weight of the second in-vehicle model, the output result of the first in-vehicle model is used as the second prediction result; when the output result of the first in-vehicle model is different from the output result of the second in-vehicle model, and the weight of the first in-vehicle model is less than or equal to the weight of the second in-vehicle model, the output result of the second in-vehicle model is used as the second prediction result.
8. The method according to claim 7, wherein During the update process of the vehicle model, the weights of the first vehicle model are updated according to whether the output result of the first vehicle model conforms to the actual operation of the user, and the weights of the second vehicle model are updated according to whether the output result of the second vehicle model conforms to the actual operation of the user.
9. The method according to any one of claims 6-8, characterized in that During the update process of the vehicle model, when a newly issued third vehicle model is received before the update is completed, the third vehicle model is used to replace the second vehicle model for the update.
10. The method according to claim 1, wherein The method further includes: Adjusting the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder.
11. A vehicle fault prediction device, characterized in that, The device includes: A training module that trains a common model to obtain a trained common model; A prediction module that inputs the vehicle data of the target vehicle into the trained common model to output a first prediction result regarding vehicle faults; inputs the vehicle data and the first prediction result into the vehicle model of the target vehicle to output a second prediction result regarding vehicle faults; and A reminder module that reminds the user of the target vehicle according to the second prediction result, wherein when the training module trains the common model, it uses the training data collected within a preset time range for training, and the weight coefficient for calculating the weight of the training data is related to the time of the training data.
12. The device according to claim 11, characterized in that The longer the time interval between the time of the training data and the current time, the smaller the weight coefficient, The shorter the time interval between the time of the training data and the current time, the larger the weight coefficient, and when the time interval between the time of the training data and the current time is less than a preset value, the weight coefficient is 1.
13. The device according to claim 11, characterized in that When the training module trains the common model, it uses the recall rate at the previous moment as the weight of the accuracy rate at the current moment, and uses the accuracy rate at the previous moment as the weight of the recall rate at the current moment to calculate the score at the current moment.
14. The device according to claim 11, characterized in that The weight of the training data is calculated according to the weight coefficient and the corruption coefficient, When the weight coefficient is greater than the corruption coefficient, the weight of the training data is the difference between the weight coefficient and the corruption coefficient, When the weight coefficient is less than or equal to the corruption coefficient, the weight of the training data is zero.
15. The device according to claim 11, wherein The device further includes: An update module that periodically updates the vehicle model of the target vehicle, During the update process of the vehicle model, the first vehicle model and the second vehicle model are kept running, the first vehicle model is the original model, and the second vehicle model is the newly issued model.
16. The device according to claim 15, characterized in that During the update process of the vehicle model, When the output result of the first vehicle-mounted model is the same as the output result of the second vehicle-mounted model, use the output result of the first vehicle-mounted model as the second prediction result; When the output result of the first vehicle-mounted model is different from the output result of the second vehicle-mounted model, and the weight of the first vehicle-mounted model is greater than the weight of the second vehicle-mounted model, use the output result of the first vehicle-mounted model as the second prediction result; When the output result of the first vehicle-mounted model is different from the output result of the second vehicle-mounted model, and the weight of the first vehicle-mounted model is less than or equal to the weight of the second vehicle-mounted model, use the output result of the second vehicle-mounted model as the second prediction result.
17. The device according to claim 15, wherein: During the update process of the vehicle-mounted model, the update module updates the weight of the first vehicle-mounted model according to whether the output result of the first vehicle-mounted model conforms to the actual operation of the user, and updates the weight of the second vehicle-mounted model according to whether the output result of the second vehicle-mounted model conforms to the actual operation of the user.
18. The device according to any one of claims 15-17, wherein: During the update process of the vehicle-mounted model, when a newly issued third vehicle-mounted model is received before the update is completed, the update module uses the third vehicle-mounted model to replace the second vehicle-mounted model for update.
19. The device according to claim 11, wherein: The reminder module adjusts the time interval and reminder probability of the next reminder according to whether the user has performed relevant operations on the reminder.
20. A vehicle maintenance system, characterized in that, The system includes: A maintenance structure that repairs the vehicle; A server that stores a trained common model. When training the common model, it is trained using training data collected from the maintenance structure within a preset time range, and the weight coefficient for calculating the weight of the training data is related to the time of the training data; A vehicle head unit system that stores a vehicle-mounted model; The server inputs the vehicle data of the target vehicle into the trained common model, outputs a first prediction result regarding vehicle faults; and sends the first prediction result to the vehicle head unit system of the target vehicle, The vehicle head unit system of the target vehicle inputs the vehicle data and the first prediction result into the vehicle-mounted model, outputs a second prediction result regarding vehicle faults, and gives a reminder to the user of the target vehicle according to the second prediction result.