Vehicle failure prediction method, vehicle, and vehicle failure prediction system

Abstract

The embodiment of the application provides a vehicle fault prediction method, a vehicle and a vehicle fault prediction system, and belongs to the technical field of intelligent networked vehicles. The method comprises the following steps: obtaining training fault data of a target vehicle; training an original fault prediction model according to the training fault data to obtain original gradient data of an updated fault prediction model; wherein the original fault prediction model is issued by a cloud and is deployed in a light weight manner; performing data disturbance processing on the original gradient data to obtain encrypted gradient data; sending the encrypted gradient data to the cloud, and receiving model increment information of the updated fault prediction model fed back by the cloud according to the encrypted gradient data; updating the updated fault prediction model according to the model increment information to obtain a target fault prediction model; and performing fault prediction on the target vehicle through the target fault prediction model. The embodiment of the application can realize joint training between the cloud and the vehicle-mounted terminal, guarantee the training effect, reduce the risk of data leakage, and improve the data transmission security.

Description

Technical Field

[0001] This application relates to the field of intelligent connected vehicle technology, and in particular to a vehicle fault prediction method, a vehicle, and a vehicle fault prediction system. Background Technology

[0002] Vehicle fault prediction has shifted from traditional rule-based prediction to fault prediction models. The performance of these models is highly dependent on the scale and diversity of the training data; effectively utilizing dispersed vehicle fault data can improve prediction accuracy. In related technologies, fault prediction models are trained in the cloud using federated learning across multiple vehicles. These trained models are then deployed on the vehicle, where they analyze received sensor data to predict faults and output the results. However, before training the model, the cloud struggles to collect large amounts of data from the vehicle, leading to insufficient accuracy in the trained model. Furthermore, while vehicle-side fault prediction results can update the cloud-based model, the communication complexity between the vehicle and cloud is high, and data transmission security is insufficient.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a vehicle fault prediction method, a vehicle, and a vehicle fault prediction system. The aim is to enable joint training between the cloud and the vehicle terminal during the fault prediction model training process, thereby improving the training effect, reducing the risk of data leakage, and enhancing the security of data transmission.

[0005] To achieve the above objectives, one aspect of this application proposes a vehicle fault prediction method, executed by an on-board terminal, wherein the on-board terminal and the cloud are connected in communication. The method includes the following steps: Acquire training fault data for the target vehicle; The preset original fault prediction model is trained based on the training fault data to obtain an updated fault prediction model and the original gradient data of the updated fault prediction model; wherein, the original fault prediction model is distributed and deployed in a lightweight manner by the cloud. The original gradient data is subjected to data perturbation processing to obtain encrypted gradient data; The encrypted gradient data is sent to the cloud, and the cloud is received as feedback of the model increment information for updating the fault prediction model based on the encrypted gradient data. The updated fault prediction model is updated based on the incremental information of the model to obtain the target fault prediction model; The target vehicle is used to predict faults using the target fault prediction model.

[0006] In some embodiments, the step of performing data perturbation processing on the original gradient data to obtain encrypted gradient data includes: The original gradient data is clipped according to preset clipping parameters to obtain the first gradient data; The first gradient data is subjected to noise processing to obtain the second gradient data; The second gradient data is compressed to obtain compressed gradient data; The compressed gradient data is anonymized to obtain the third gradient data; The third gradient data is encrypted to obtain the encrypted gradient data.

[0007] In some embodiments, training a preset original fault prediction model based on the training fault data to obtain an updated fault prediction model and the original gradient data of the updated fault prediction model includes: Obtain the current operating status and current remaining energy of the target vehicle; The target vehicle is trained and evaluated based on the current operating status and the current remaining energy to obtain the training and evaluation results. If the training evaluation result indicates that the target vehicle meets the training conditions, the original fault prediction model is trained according to the training fault data and the preset training configuration parameters to obtain an updated fault prediction model. Obtain the original gradient data of the updated fault prediction model.

[0008] In some embodiments, after obtaining the original gradient data of the updated fault prediction model, the method further includes: Obtain the first prediction accuracy of the original fault prediction model and the second prediction accuracy of the updated fault prediction model; Obtain the amount of training fault data; The accuracy improvement value is determined based on the first prediction accuracy and the second prediction accuracy; The accuracy improvement value, the amount of data, and the training configuration parameters are sent to the cloud, and resource values ​​sent by the cloud based on the accuracy improvement value, the amount of data, and the training configuration parameters are received; wherein, the resource values ​​are used for resource exchange on the resource exchange platform of the target vehicle.

[0009] To achieve the above objectives, another aspect of this application proposes a vehicle fault prediction method, executed by a cloud, wherein the cloud is communicatively connected to at least one in-vehicle terminal, the method comprising: The original fault prediction model is sent to the vehicle terminal; The system receives encrypted gradient data fed back by at least one vehicle-mounted terminal based on the original fault prediction model; wherein the encrypted gradient data is obtained by the vehicle-mounted terminal training a preset original fault prediction model based on collected training fault data to obtain original gradient data for updating the fault prediction model, and then performing data perturbation processing on the original gradient data. The original fault prediction model is optimized based on at least one of the encrypted gradient data to obtain the current fault prediction model. Determine model increment information based on the current fault prediction model and the original fault prediction model; The model increment information is sent to the vehicle terminal so that the vehicle terminal updates the updated fault prediction model according to the model increment information to obtain the target fault prediction model, and performs fault prediction on the target vehicle through the target fault prediction model.

[0010] In some embodiments, optimizing the original fault prediction model based on at least one of the encrypted gradient data to obtain the current fault prediction model includes: Masking at least one of the encrypted gradient data yields at least one model gradient value. Anomaly assessment is performed on the gradient values ​​of the model to obtain anomaly assessment data; Based on the anomaly assessment data, reliable gradient values ​​are selected from the model gradient values; The original fault prediction model is optimized based on the reliable gradient value to obtain the current fault prediction model.

[0011] In some embodiments, the step of performing anomaly evaluation on the model gradient values ​​to obtain anomaly evaluation data includes: Calculate the distance between any two gradient values ​​of the model to obtain the gradient distance; Selected gradient values ​​are filtered from the model gradient values ​​based on a preset quantity and the gradient distance; The gradient distances are concatenated based on the selected gradient values ​​to obtain the anomaly assessment data.

[0012] In some embodiments, after sending the model incremental information to the vehicle terminal, the method further includes: Receive the accuracy improvement, data volume, and training configuration parameters sent by the vehicle terminal; Rarity assessment is performed on the updated fault prediction model to obtain rarity assessment data; The contribution of the vehicle terminal is evaluated based on the accuracy improvement, the amount of data, the training configuration parameters, and the rarity assessment data to obtain the current contribution. The vehicle terminal is allocated resources based on its current contribution level and a preset resource allocation table.

[0013] To achieve the above objectives, another aspect of this application provides a vehicle comprising: an on-board terminal, a vehicle body, and at least one sensor. The on-board terminal includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method executed by the on-board terminal described above.

[0014] To achieve the above objectives, another aspect of this application provides a vehicle fault prediction system, which includes at least one vehicle-mounted terminal and a cloud. The vehicle-mounted terminal is equipped with a first computer program, and the cloud is equipped with a second computer program. When the first computer program is executed by a processor, it implements the method executed by the vehicle-mounted terminal, and when the second computer program is executed by a processor, it implements the method executed by the cloud.

[0015] The embodiments of this application include at least the following beneficial effects: This application provides a vehicle fault prediction method, a vehicle, and a vehicle fault prediction system. This solution involves lightweight deployment of an original fault prediction model distributed from the cloud on an in-vehicle terminal, and training fault data of the target vehicle to train the original fault prediction model to obtain an updated fault prediction model. The gradient data of the updated fault prediction model is encrypted and transmitted to the cloud. The cloud, in conjunction with gradient data from multiple in-vehicle terminals, completes global optimization of the original fault prediction model and outputs model increment information. The in-vehicle terminal updates its local fault prediction model based on the model increment information, outputting a target fault prediction model with higher prediction accuracy. The target fault prediction model is then used to predict the fault of the target vehicle. Therefore, during the fault prediction model training process, the cloud collaborates with multiple in-vehicle terminals, and only the in-vehicle terminal needs to initially train the fault prediction model locally. The encrypted gradient data uploaded by the cloud and multiple in-vehicle terminals further optimizes the fault prediction model, improving the training effect of the fault prediction model, reducing the risk of data leakage, and enhancing the security of data transmission. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the vehicle fault prediction method provided in the embodiments of this application; Figure 2 This is a detailed structural diagram of the vehicle fault prediction method provided in the embodiments of this application; Figure 3 This is a flowchart of the vehicle fault prediction method provided in the embodiments of this application; Figure 4 yes Figure 3 The flowchart of step S302 in the text; Figure 5 This is a flowchart of a vehicle fault prediction method provided in another embodiment of this application; Figure 6 yes Figure 3 The flowchart of step S303 in the process; Figure 7 This is a flowchart illustrating privacy protection in the vehicle fault prediction method provided in this application embodiment; Figure 8 This is a flowchart of a vehicle fault prediction method provided in another embodiment of this application; Figure 9 yes Figure 8 The flowchart of step S803 in the process; Figure 10 yes Figure 9 The flowchart of step S902 in the text; Figure 11 This is a flowchart of a vehicle fault prediction method provided in another embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0018] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0019] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0022] 1) Federated Machine Learning, also known as federated learning, is a distributed machine learning method that enables multi-party joint modeling through encrypted parameter exchange, ensuring that training can be completed without the data leaving the local device or institution. It is mainly used in fields such as financial services, logistics, and edge computing.

[0023] 2) A global model is a shared machine learning model that is collaboratively trained by multiple distributed devices or servers without centralizing the raw data. Each participant (e.g., a smartphone, IoT device, or local server) trains a local model on its own data, rather than sending data to a central server. These local updates are then aggregated to create a global model, which serves as a unified representation of the knowledge learned from all participants.

[0024] 3) Model gradient: In deep learning and machine learning, the gradient is a core tool for optimizing model parameters. It is a vector composed of the partial derivatives of the loss function with respect to each parameter, pointing in the direction of the fastest increase in the function value. Moving along the negative gradient direction will reduce the loss function value the fastest.

[0025] 4) Differential Privacy (DP): This involves injecting random noise into statistical results or model parameters to ensure that attackers cannot determine whether a particular individual's data exists in the dataset. The core of adding noise is the noise mechanism, with common methods including Laplace's noise reduction and Gaussian noise reduction.

[0026] 5) Homomorphic Encryption (HE) is a special encryption technique that allows direct computation on ciphertext without decryption. Its core characteristic is that computations performed on ciphertext, after decryption, yield the same result as the same operation performed on plaintext. This technique achieves "computable but not visible" data and is widely used in fields such as cloud computing, blockchain, and privacy computing.

[0027] 6) Secure aggregation protocols are a technology used to protect user privacy and are widely applied in federated learning. The core objective is to ensure that servers cannot obtain specific information about individual users when aggregating user data, thereby achieving privacy protection.

[0028] 7) Byzantine Fault Tolerance (BFT) is an important mechanism for achieving fault tolerance in distributed computing systems, designed to ensure that the system can still operate normally in the presence of malicious or faulty nodes.

[0029] The widespread adoption of intelligent connected vehicles is driving a shift in vehicle fault prediction technology from traditional rule-based prediction to machine learning-based intelligent prediction. The performance of machine learning models is highly dependent on the scale and diversity of training data. However, vehicle fault data is scattered across millions of vehicles, making the effective utilization of this data a key challenge for the industry. Among related technologies, federated learning, as an emerging distributed machine learning paradigm, offers a new technical path to resolving the conflict between data silos and privacy protection. Currently, federated learning-based vehicle fault prediction primarily involves multiple automakers uploading fault data to train a fault prediction model. This trained model is then deployed to in-vehicle terminals, where the model predicts faults in the target vehicle and outputs the results. During the joint training process, data privacy restrictions prevent the combined use of data uploaded from multiple in-vehicle terminals, impacting prediction accuracy. Furthermore, data transmission between in-vehicle terminals and the cloud poses security risks.

[0030] In view of this, this application provides a vehicle fault prediction method, a vehicle, and a vehicle fault prediction system. The method involves lightweight deployment of an original fault prediction model distributed from the cloud on an in-vehicle terminal, and training fault data of the target vehicle to train the original fault prediction model to obtain an updated fault prediction model. Then, the gradient data of the updated fault prediction model is encrypted and transmitted to the cloud. The cloud, in conjunction with gradient data from multiple in-vehicle terminals, performs global optimization of the original fault prediction model and outputs model increment information. The in-vehicle terminal updates its local fault prediction model based on the model increment information, outputting a target fault prediction model with higher prediction accuracy. Fault prediction of the target vehicle is then achieved through this target fault prediction model. Therefore, during the fault prediction model training process, the cloud collaborates with multiple in-vehicle terminals. Only the in-vehicle terminal needs to initially train the fault prediction model locally, and the encrypted gradient data uploaded by the cloud and multiple in-vehicle terminals further optimizes the fault prediction model, improving the training effect of the fault prediction model, reducing the risk of data leakage, and enhancing the security of data transmission.

[0031] The vehicle fault prediction method provided in this application relates to the field of intelligent vehicle networking technology. The vehicle fault prediction method provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the vehicle fault prediction method, but is not limited to the above forms.

[0032] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0033] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0034] Please refer to Figure 1 , Figure 1The system framework diagram of the vehicle fault prediction model provided in the embodiment of this application is shown, which includes at least one vehicle terminal and a cloud, and a communication connection between the vehicle terminal and the cloud.

[0035] The vehicle-mounted terminal is used to collect training fault data of the target vehicle and, in conjunction with the cloud, train a fault prediction model. The trained model is then used to predict faults in the target vehicle. The vehicle-mounted terminal can take various forms, including desktop computers, laptops, tablets, mobile phones, in-vehicle terminals, home theater terminals, and dedicated terminals, all connected to the target vehicle. Furthermore, the vehicle-mounted terminal can be a single device or a collection of multiple devices.

[0036] The cloud refers to a cluster of computing devices, such as a data center or server cluster, used to train a fault prediction model in conjunction with at least one in-vehicle terminal. The cloud can be a single server or a server cluster consisting of multiple servers.

[0037] Please refer to Figure 2 The vehicle-mounted terminal includes: a data acquisition layer, a local training layer, a privacy protection layer, and a communication layer. The data acquisition layer includes a buffer, a data acquisition unit, and a recorder. The buffer uses encrypted storage to store high-frequency sensor data triggered by a fault. The data acquisition unit acquires data from sensors, including multi-source data such as vibration, acoustics, and CAN bus data. The recorder records voice or text information sent by the vehicle-mounted terminal. The local training layer includes: a lightweight training engine, a gradient calculator, a differential privacy module, and a model parameter storage. The lightweight training engine is based on TensorFlow Lite or PyTorch Mobile and supports INT8 quantization training. The gradient calculator calculates the gradient data by comparing the parameters of the local fault prediction model and the global fault prediction model. The differential privacy module implements a differential privacy mechanism, and the model parameter storage stores the intermediate states of the current global fault prediction model and the locally trained fault prediction model. The privacy protection layer includes a homomorphic encryption module, a gradient compression layer, an anonymization processor, and a hardware security module. The homomorphic encryption module uses the Pailler or CKKS scheme to encrypt gradients, and the gradient compressor uses Top-K sparsity or random sparsity to reduce communication overhead. The anonymization processor is used to anonymize vehicle identity information. The hardware security module provides a trusted execution environment to protect keys. The communication layer includes an upload scheduler, a breakpoint resume module, an OTA receiver, and a communication unit. The upload scheduler detects the charging status of the target vehicle and the communication status between the on-board terminal and the cloud, intelligently scheduling the upload tasks of the on-board terminal. The breakpoint resume module supports the interruption and resumption of upload tasks, and the OTA receiver receives fault prediction models from the cloud.

[0038] The cloud-based system includes a federated learning service platform, a model management service layer, and an incentive service layer. The federated learning service platform comprises a participant manager, a gradient receiver and acceptor, a secure aggregation engine, a global model trainer, and an anomaly detector. The participant manager manages the list and status of vehicles participating in federated learning. The gradient receiver and acceptor receives encrypted gradients and verifies their integrity. The secure aggregation engine uses the SecAgg protocol for secure aggregation, and the protocol used by the secure aggregation engine is not limited. The global model trainer updates the global fault prediction model and can use either the FedAvg or FedProx algorithm; this embodiment does not limit the training algorithm. The anomaly detector uses the Krum algorithm to detect and filter Byzantine attacks. The model management layer includes a model repository, an incremental update generator, an OTA distribution scheduler, and a model quality evaluator. The model repository manages different versions of the global fault prediction model. The incremental update generator generates model parameter differences to obtain incremental model information, reducing the amount of data distributed. The OTA distribution scheduler supports A / B testing and canary releases, and the model quality evaluator evaluates the performance and security of the updated fault prediction model. The incentive service platform includes: a contribution calculation engine, a resource account system, a rights exchange interface, and an audit log service. The contribution calculation engine is used to quantify the data contributions provided by users from multiple dimensions. The resource account system is used to manage the resource values ​​and historical values ​​of the vehicle terminal. The rights exchange interface is used to connect to the after-sales service system and support resource exchange. The audit log service is used to record all operations of the vehicle terminal and support subsequent auditing and traceability.

[0039] In summary, the specific structural settings between the vehicle terminal and the cloud aim to achieve data privacy protection, improve communication efficiency between the vehicle terminal and the cloud, enhance fault prediction accuracy, increase the participation rate of the vehicle terminal in the fault prediction model, reduce gradient attack rate, and reduce the convergence rounds of the fault prediction model.

[0040] Please refer to Figure 3 , Figure 3 This is an optional flowchart of the vehicle fault prediction method provided in the embodiments of this application. Figure 3 The method may include, but is not limited to, steps S301 to S306.

[0041] Step S301: Obtain training fault data for the target vehicle; Step S302: Train the preset original fault prediction model based on the training fault data to obtain the updated fault prediction model and the original gradient data of the updated fault prediction model; wherein, the original fault prediction model is distributed from the cloud and deployed in a lightweight manner. Step S303: Perform data perturbation processing on the original gradient data to obtain encrypted gradient data; Step S304: Send the encrypted gradient data to the cloud and receive the model increment information updated by the cloud based on the encrypted gradient data to update the fault prediction model; Step S305: Update the fault prediction model according to the model incremental information to obtain the target fault prediction model; Step S306: Perform fault prediction on the target vehicle using the target fault prediction model.

[0042] Steps S301 to S306 of this embodiment involve first deploying a lightweight original fault prediction model at the vehicle terminal, then training the original fault prediction model locally using training fault data collected by the vehicle terminal to obtain an updated fault prediction model and its original gradient data. Encrypted gradient data is obtained by perturbing the original gradient data and then uploaded to the cloud. The cloud then combines the encrypted gradient data uploaded by multiple vehicle terminals to complete the original fault prediction model and distributes incremental model information to each vehicle terminal. Therefore, the fault prediction model training process implements a federated learning training mode, ensures data privacy, increases the number of vehicle terminals participating in the training, and improves model training effectiveness. Furthermore, only encrypted gradient data is transmitted during gradient data transmission, enhancing data security. Thus, a fault prediction model that ensures both effective training and secure data transmission can improve the accuracy of fault prediction.

[0043] In step S301 of some embodiments, the training fault data is sensor data collected in advance when the target vehicle malfunctions, specifically including vibration, acoustic, and CAN interface data, etc., and the training fault data is encrypted and stored in a local database in advance. It should be noted that the collection of training fault data is only executed upon a fault trigger command or sensor activation. The fault trigger command can be directly input into the vehicle terminal or generated via voice; this embodiment does not limit the generation method of the fault trigger command. After collecting sensor data such as vibration, acoustic, and CAN interface data, the collected sensor data is preprocessed, including noise reduction, standardization, and feature extraction, and then the preprocessed sensor data is encrypted and stored in the local database. When the vehicle terminal completes the deployment of the original fault prediction model and needs to train the original fault prediction model, sensor data is extracted from the local database and defined as training fault data.

[0044] As previously disclosed, before acquiring training fault data, the vehicle terminal receives the original fault prediction model sent from the cloud and initialized. It should be noted that the global fault prediction model sent from the cloud is defined as G_model, and the original fault prediction model is obtained by initializing the global fault prediction model. The initialization specifically includes: (1) verifying the signature of the global fault prediction model to ensure the integrity of the global fault prediction model; (2) creating a copy of the global fault prediction model to obtain the original fault prediction model, defined as L_model=copy(G_model); (3) loading the training configuration parameters of the vehicle terminal from the cloud. The cloud packages the original fault prediction model and the training configuration parameters to obtain a model data package, and sends the model data package to the vehicle terminal with communication connection through the OTA secure distribution model. After receiving the model data package, the vehicle terminal deploys the original fault prediction model in a lightweight manner.

[0045] Please see Figure 4 In some embodiments, step S302 may include, but is not limited to, steps S401 to S404: Step S401: Obtain the current operating status and current remaining energy of the target vehicle; Step S402: Based on the current operating status and the current remaining energy, perform training evaluation on the target vehicle to obtain the training evaluation results; Step S403: If the training evaluation result indicates that the target vehicle meets the training conditions, the original fault prediction model is trained according to the training fault data and the preset training configuration parameters to obtain the updated fault prediction model. Step S404: Obtain the original gradient data for updating the fault prediction model.

[0046] In step S401 of some embodiments, the current operating state includes a start-up state, an off state, and a charging state, and the current remaining energy represents the remaining battery power or remaining fuel of the target vehicle. The current operating state and the current remaining energy are used to assess whether the on-board terminal meets the training conditions.

[0047] In step S402 of some embodiments, the training evaluation mainly assesses whether the on-board terminal corresponding to the target vehicle can start the training of the original fault prediction model. Specifically, the current operating state is compared with the off-state and charging states to obtain first comparison information, and the current remaining energy is compared with a preset energy threshold to obtain second comparison information. The training evaluation result is determined based on the first comparison information and the second comparison information. If the first comparison information indicates that the current operating state is off-state or charging state, and the second comparison information indicates that the current remaining energy is greater than the preset energy threshold, the training evaluation result indicates that the target vehicle meets the training conditions; otherwise, if the first comparison information indicates that the current operating state is on-state, or the second comparison information indicates that the current remaining energy is less than or equal to the preset energy threshold, the training evaluation result indicates that the target vehicle does not meet the training conditions and cannot support the training of the original fault prediction model.

[0048] For example, if the preset energy threshold is 30% and the current remaining energy is the current remaining battery power, if the target vehicle is in a turned-off state or the target vehicle is in a charging state and the current remaining battery power is greater than 30%, the training of the original fault prediction model will be started.

[0049] In step S403 of some embodiments, the training fault data is divided into a training set and a validation set according to a preset ratio, and the original fault prediction model is trained according to the training configuration parameters, the training set, and the validation set, and an updated fault prediction model is output. It should be noted that the training configuration parameters include: training epochs, batch size, and learning rate. The original fault prediction model is trained according to the batch size and training epochs, and the original fault prediction model is converged according to the learning rate, and an updated fault prediction model is output.

[0050] In step S404 of some embodiments, the gradient data of the updated fault prediction model is obtained as the original gradient data. It should be noted that the original gradient data of the updated fault prediction model is a vector representing the model update information from the original fault prediction model to the updated fault prediction model. Specifically, it is a vector composed of the partial derivatives of the loss function with respect to the parameters of the original fault prediction model, pointing in the direction of the fastest growth of the loss function. This is equivalent to telling the original fault prediction model the direction and amount of parameter adjustment, so that a fault prediction model matching the updated fault prediction model can be generated in the cloud using the original gradient data.

[0051] In steps S401 to S404 of this embodiment, the current operating status and remaining energy of the target vehicle are collected. The current operating status and remaining energy are used to assess whether the target vehicle meets the training conditions. Only if the conditions are met is the training of the original fault prediction model initiated, and the original gradient data for updating the fault prediction model after training collected. Therefore, training the model only after the training conditions are met reduces excessive energy loss of the target vehicle and minimizes the impact of model training on the use of the target vehicle.

[0052] After step S302 in some embodiments, the vehicle fault prediction method further includes: evaluating the updated fault prediction model, uploading the evaluation results and training data of the updated fault prediction model to the cloud, and receiving resource values ​​fed back from the cloud.

[0053] Traditional federated learning requires multiple in-vehicle terminals to participate in training global and local models. However, users are generally unwilling to participate in model training, leading to insufficient prediction accuracy in the trained fault prediction models. There is a lack of mature solutions for quantifying user data contributions and providing reasonable incentives, necessitating the establishment of a fair and transparent contribution calculation model. Therefore, this embodiment sets up a contribution quantification and resource incentive mechanism to transform users into model training participants, improving model training effectiveness. It should be noted that traditional user participation rates are typically 10%-20%, with data coverage of approximately 60% and user satisfaction of 60%. The resource incentive mechanism set up in this embodiment can increase user participation to 60%-80%, with data coverage exceeding 98% and user satisfaction reaching 80%. Therefore, adopting the resource incentive mechanism of this embodiment significantly improves user participation, data coverage, and user satisfaction.

[0054] Please see Figure 5 In some embodiments, the vehicle fault prediction method may also include, but is not limited to, steps S501 to S504: Step S501: Obtain the first prediction accuracy of the original fault prediction model and the second prediction accuracy of the updated fault prediction model. Step S502: Obtain the amount of training fault data; Step S503: Determine the accuracy improvement value based on the first prediction accuracy and the second prediction accuracy; Step S504: Send the accuracy improvement value, data volume and training configuration parameters to the cloud, and receive the resource value sent by the cloud based on the accuracy improvement value, data volume and training configuration parameters; wherein, the resource value is used for resource exchange on the resource exchange platform of the target vehicle.

[0055] In step S501 of some embodiments, the first prediction accuracy of the original fault prediction model is determined by inputting the validation set into the original fault prediction model, and the second prediction accuracy of the updated fault prediction model is also determined by inputting the validation set.

[0056] It should be noted that the second prediction accuracy will be uploaded to the cloud. The cloud will then optimize the global fault prediction model based on each second prediction accuracy and encrypted gradient data. The second prediction accuracy helps the cloud determine the data quality or model convergence of each vehicle terminal, so as to train a fault prediction model with stronger prediction accuracy.

[0057] In step S502 of some embodiments, the data volume is also referred to as the effective sample ratio in the training fault data. Specifically, the number of effective samples and the number of benchmark samples are obtained, and the effective sample ratio is determined by the ratio between the number of effective samples and the number of benchmark samples.

[0058] In step S503 of some embodiments, the difference between the second prediction accuracy and the first prediction accuracy is obtained to determine the accuracy improvement value. The training effect of the vehicle terminal can be determined by the accuracy improvement value.

[0059] In step S504 of some embodiments, the accuracy improvement value, data volume, and training configuration parameters are sent to the cloud. The cloud performs a category rarity assessment on the updated fault prediction model to obtain rarity assessment data. Then, based on the accuracy improvement value, data volume, training rounds and training time in the training configuration parameters, and the rarity assessment data, the contribution of the vehicle terminal is quantified to obtain the data contribution. Based on the data contribution, resource values ​​are distributed to the vehicle terminal. The resource value distribution mode incentivizes the vehicle terminal to participate in model training, thereby improving the model training effect and thus improving the accuracy of vehicle fault prediction.

[0060] It should be noted that the resource value can be points earned by connecting the vehicle terminal to the resource exchange platform, or it can be tokens, insurance discounts, extended warranty benefits, etc. This embodiment does not limit the specific form of the resource value. If the resource value is points, the resource exchange platform is a points mall connected to the vehicle terminal, and the redeemable benefits are service benefits of the target vehicle, specifically including VIP services, extended warranty, free maintenance, priority appointments, discount services, basic points, and ordinary coupons.

[0061] In steps S501 to S504 of this embodiment, the vehicle terminal uploads the accuracy improvement of the updated fault prediction model, the amount of training fault data, and the training configuration parameters to the cloud. The cloud can quantify the contribution of each vehicle terminal to the fault prediction model training, provide resources to complete the rights exchange, set up an incentive mechanism for model training, increase the participation rate of vehicle terminals in model training, and thus improve the prediction accuracy of the model.

[0062] In step S303 of some embodiments, data perturbation processing is a privacy protection measure, specifically including data noise addition and data encryption. Traditionally, differential privacy noise addition is used for data privacy protection; however, excessive noise addition can affect model convergence, while insufficient noise addition cannot effectively protect privacy, requiring a precise privacy budget allocation strategy. Therefore, this embodiment employs a dual protection approach of differential privacy noise addition and homomorphic encryption to prevent gradient propagation attacks.

[0063] Please see Figure 6 In some embodiments, step S303 may include, but is not limited to, steps S601 to S605: Step S601: The original gradient data is clipped according to the preset clipping parameters to obtain the first gradient data; Step S602: Noise is added to the first gradient data to obtain the second gradient data; Step S603: Compress the second gradient data to obtain compressed gradient data; Step S604: Anonymize the compressed gradient data to obtain the third gradient data; Step S605: Encrypt the third gradient data to obtain encrypted gradient data.

[0064] In step S601 of some embodiments, as disclosed above, the original gradient data is the model parameter gradient update information of the original fault prediction model. To save the amount of gradient data transmission, the original gradient data is pruned according to pruning parameters, and the pruned first gradient data is output. The first gradient data can characterize the specific model parameter information of the updated fault prediction model. It should be noted that the pruning parameters include: a pruning threshold and a pruning factor, and the pruning threshold is set according to the norm of the original gradient data.

[0065] Specifically, please refer to Figure 7The original gradient data is defined as `gradients`, and its norm is calculated as `norm = ||gradients||_2`. The clipping threshold is determined as `C = clip_norm`, and the clipping factor is calculated as `factor = min(1, C / norm)`. Finally, based on the clipping factor and the original gradient data, the clipped gradient is determined as `clipped = gradients * factor`, and the clipped gradient in the original gradient data is removed to obtain the first gradient data.

[0066] In step S602 of some embodiments, this embodiment uses differential privacy noise addition and determines the noise data to be added to the first gradient data. Specifically, the noise data is determined based on the first gradient data, privacy budget, privacy relaxation parameter and gradient pruning norm. The noise data satisfies a preset noise function, which is a function that is provided by an actual library (such as TensorFlowPrivacy or Opacus) to automatically calculate the noise multiplier. The second gradient data is obtained by adding the noise data to the first gradient data.

[0067] In step S603 of some embodiments, the gradient data for the next day is compressed to save data transmission volume. Please refer to [link / reference needed]. Figure 7 The compression strategy is as follows: (1) Flatten the second gradient data into a one-dimensional vector to obtain the gradient vector; (2) Sort the gradient vector by absolute value and retain the first K% of the elements; (3) Set the unretained elements to zero; (4) Record the index and value of the non-zero elements to obtain the compressed gradient data.

[0068] In step S604 of some embodiments, anonymization involves processing the user identity information involved in the compressed gradient data to anonymize it. This anonymization process includes replacing the VIN code with a temporary ID, changing the ID for each training round, and preventing the tracking of historical records. The resulting third gradient data cannot obtain the user identity information from the vehicle terminal, thus reducing the risk of user privacy information leakage.

[0069] In step S605 of some embodiments, this embodiment uses homomorphic encryption to encrypt the third gradient data. Please refer to... Figure 7 The encryption process is as follows: (1) The third gradient data is encrypted using a preset public key, and the preset public key is the Parller public key. This embodiment does not restrict the preset public key. (2) The encrypted gradient data can be added in the ciphertext field. (3) The data is aggregated in the cloud and then decrypted by the vehicle manufacturer's private key of the vehicle terminal. Therefore, by setting encryption, the cloud can only see the encrypted gradient data and cannot obtain the gradient data of a single vehicle terminal. It can only obtain the aggregated plaintext result.

[0070] In steps S601 to S605 of this embodiment, the original gradient data is pruned, noise-added, compressed, anonymized, and encrypted. The noise addition employs differential privacy noise, adding carefully designed random noise to the gradients to hide the impact of any single training failure data point within the noise. This ensures that even if an attacker obtains the noisy gradient data, they cannot infer the training failure data, thus improving data privacy and security. Simultaneously, homomorphic encryption generates encrypted gradient data, further protecting the security of the gradient data during transmission.

[0071] Please see Figure 8 In addition, embodiments of this application also disclose a vehicle fault prediction method executed in the cloud. Figure 8 This may include, but is not limited to, steps S801 to S805: Step S801: Send the original fault prediction model to the vehicle terminal; Step S802: Receive encrypted gradient data fed back by at least one vehicle terminal based on the original fault prediction model; wherein, the encrypted gradient data is obtained by the vehicle terminal training the preset original fault prediction model based on the collected training fault data to obtain the original gradient data for updating the fault prediction model, and performing data perturbation processing on the original gradient data. Step S803: Optimize the original fault prediction model based on at least one encrypted gradient data to obtain the current fault prediction model; Step S804: Determine the model increment information based on the current fault prediction model and the original fault prediction model; Step S805: The model increment information is sent to the vehicle terminal so that the vehicle terminal can update the updated fault prediction model according to the model increment information to obtain the target fault prediction model, and perform fault prediction on the target vehicle through the target fault prediction model.

[0072] In step S801 of some embodiments, as disclosed above, a global fault prediction model is set up in the cloud, and an original fault prediction model is obtained after initializing the global fault prediction model. Then, the original fault prediction model is sent to each connected vehicle terminal. It should be noted that the global fault prediction model is trained by the cloud based on historical maintenance data. In order to further improve fault prediction, at least one vehicle terminal is used to further optimize the global fault prediction model.

[0073] In step S802 of some embodiments, each vehicle terminal trains the original fault prediction model into an updated fault prediction model and uploads the encrypted gradient data of the updated fault prediction model. It should be noted that the training process of training the original fault prediction model into the updated fault prediction model is as described above and will not be repeated here. Each vehicle terminal, in conjunction with its own collected training fault data, trains the original fault prediction model into the updated fault prediction model; therefore, the encrypted gradient data uploaded by each vehicle terminal is different.

[0074] In step S803 of some embodiments, the encrypted gradient data is a masked gradient. Since the masks cancel each other out, at least one encrypted gradient data is aggregated to obtain the sum of the true gradients. The original fault prediction model is then trained specifically using the sum of the true gradients to obtain a more accurate current fault prediction model. It should be noted that the current fault prediction model is a global fault prediction model.

[0075] In step S804 of some embodiments, the model increment information represents the model parameter difference between the current fault prediction model and the original fault prediction model. Specifically, the model parameter difference between the original fault prediction model and the current fault prediction model is calculated, the model parameter difference is quantized and compressed, the compressed model parameter difference is used to generate a binary difference packet, and the difference packet is signed to obtain the model increment information.

[0076] It should be noted that the incremental information of the model is 5-10% of the current fault prediction model, saving the amount of data to be distributed.

[0077] In step S805 of some embodiments, the incremental model information is first sent to a subset of selected vehicle terminals, and model performance and stability information are obtained from the selected vehicle terminals. If the model performance and stability information indicate that the model is normal, the incremental model information is further sent to randomly selected vehicle terminals according to region / vehicle type, and the model feedback information from the vehicle terminals is collected. If the model feedback information within a preset number of days indicates that the model is normal, the incremental model information is then sent to all vehicle terminals. Therefore, this embodiment sets up a gray-scale release of incremental model information, which saves data transmission volume on the one hand, and ensures that the model can run stably and accurately on vehicle terminals on the other hand, thus improving the availability of the model.

[0078] In steps S801 to S805 of this embodiment, the cloud first sends the original fault prediction model to the vehicle terminal, then receives and completes global training of the original fault prediction model based on encrypted gradient data uploaded by at least one vehicle terminal to output the current fault prediction model. Finally, the model increment information between the current fault prediction model and the original fault prediction model is sent to the vehicle terminal. Therefore, federated learning training of the fault prediction model is achieved between the vehicle terminal and the cloud, and the vehicle terminal trains locally and uploads encrypted gradient data during the training process, improving data privacy and enhancing data transmission security.

[0079] Please see Figure 9 In some embodiments, step S803 includes, but is not limited to, steps S901 to S904: Step S901: Mask at least one encrypted gradient data to obtain at least one model gradient value; Step S902: Perform anomaly evaluation on the model gradient values ​​to obtain anomaly evaluation data; Step S903: Select reliable gradient values ​​from the model gradient values ​​based on the anomaly evaluation data; Step S904: Optimize the original fault prediction model based on the confidence gradient value to obtain the current fault prediction model.

[0080] In step S901 of some embodiments, the encrypted gradient data is generated by each vehicle terminal using a shared key to create a pseudo-random mask, which is then added to the original gradient data to generate masked encrypted gradient data. It should be noted that each vehicle terminal generates a temporary key pair and exchanges public keys with other participating vehicle terminals to establish a pairwise shared key, which is then used to generate the pseudo-random mask. When the cloud receives the encrypted gradient data, since the masks cancel each other out, the aggregation result is the sum of the true gradients. Therefore, at least one encrypted gradient data is aggregated to obtain the true model gradient value.

[0081] In step S902 of some embodiments, anomaly assessment is used to evaluate the degree of anomaly in the model gradient values ​​to determine anomaly assessment data. Specifically, this embodiment uses the Byzantine fault-tolerant algorithm to assess the degree of anomaly in the model gradient values. The Byzantine fault-tolerant algorithm is a malicious node identification algorithm based on Euclidean distance, used to identify whether the model gradient values ​​sent by the vehicle-mounted terminal connected to the cloud are abnormal gradient values, so as to reduce malicious behavior of some vehicle-mounted terminals.

[0082] In steps S903 and S904 of some embodiments, the model gradient value where the abnormal evaluation data is less than a preset threshold is taken as a reliable gradient value. The original fault prediction model is then jointly optimized based on at least one reliable gradient value to output a more accurate current fault prediction model. It should be noted that each model gradient value represents the model parameter update information of the updated fault prediction model of the vehicle terminal relative to the original fault prediction model. The weights of each vehicle terminal are pre-configured, and the global fault prediction model is updated based on the weights and model gradient values ​​to output the current fault prediction model.

[0083] In steps S901 to S904 of this embodiment, at least one model gradient value is obtained by aggregating encrypted gradient data to remove the mask, and a reliable gradient value is selected from the model gradient value. The original fault prediction model is updated with at least one reliable gradient value to train a current fault prediction model with more accurate fault prediction.

[0084] Please see Figure 10 In some embodiments, step S902 may include, but is not limited to, steps S1001 to S1003: Step S1001: Calculate the distance between any two model gradient values ​​to obtain the gradient distance; Step S1002: Select the gradient values ​​from the model gradient values ​​according to the preset quantity and gradient distance; Step S1003: The gradient distances are concatenated according to the selected gradient values ​​to obtain anomaly assessment data.

[0085] In step S1001 of some embodiments, any model gradient value is selected as the selected gradient value, and the gradient distance between the selected gradient value and other model gradient values ​​is calculated to obtain the gradient distance between the selected gradient value and other model gradient values. The gradient distance represents the degree of difference between the selected gradient value and other model gradient values.

[0086] In step S1002 of some embodiments, this embodiment selects a preset number of model gradient values ​​that are closest to the gradient values ​​as the selected gradient values ​​according to the gradient distance. It should be noted that the preset number can be nf-2, where n is the number of model gradient values, f is the preset number of tolerable additional nodes, and f < n / 3.

[0087] In step S1003 of some embodiments, the gradient distances corresponding to the selected gradient values ​​are summed to obtain the gradient sum, and the gradient sum is used as the anomaly score of the selected gradient value to obtain anomaly evaluation data. Therefore, the selected gradient value with the lowest anomaly score is taken as the reliable gradient value.

[0088] In steps S1001 to S1003 of this embodiment, the distance between model gradient values ​​is used to determine abnormal evaluation data of model gradient values. This can filter out model gradient values ​​with small differences as reliable gradient values, making the selection of model gradient values ​​simpler and more accurate.

[0089] Please refer to Figure 11 In some embodiments, after step S805, the vehicle fault prediction method may also include, but is not limited to, steps S1101 to S1104: Step S1101: Receive the accuracy improvement, data volume, and training configuration parameters sent by the vehicle terminal; Step S1102: Perform a rarity assessment on the updated fault prediction model to obtain rarity assessment data; Step S1103: Evaluate the contribution of the vehicle terminal based on the accuracy improvement, data volume, training configuration parameters, and rarity assessment data to obtain the current contribution. Step S1104: Allocate resources to the vehicle terminal according to the current contribution level and the preset resource allocation table.

[0090] In step S1101 of some embodiments, as disclosed above, the accuracy improvement, data volume, and training configuration parameters are determined by the vehicle terminal, and will not be repeated here.

[0091] In step S1102 of some embodiments, the rarity assessment involves classifying the updated fault prediction model to obtain model categories, calculating the proportion of each model category, and determining the rarity assessment data based on the proportion. For example, if the number of participating vehicle terminals is 5000, the proportion of model category A is determined to be 1 / 50, and the rarity assessment data is determined to be 2%. If the proportion of model category B is 1 / 4, the rarity assessment data is determined to be 25%.

[0092] In step S1103 of some embodiments, the contribution evaluation specifically determines the score corresponding to the accuracy improvement, data volume, training configuration parameters and rarity evaluation data through a preset score table, and then calculates the current contribution based on the score and preset weight.

[0093] Specifically, the first score for the accuracy improvement is defined as DataQuality, the second score for the amount of data is DataQuantity, the third score for the training configuration parameters is ParticipationFreq, and the fourth score for the rarity evaluation data is DataRarity. If the corresponding weights are 0.3, 0.25, 0.2, and 0.25, the current contribution is defined as ContributionScore, and ContributionScore = 0.30 × DataQuality + 0.25 × DataQuantity + 0.20 × ParticipationFreq + 0.25 × DataRarity.

[0094] In step S1104 of some embodiments, the resource allocation table set in this embodiment is shown in Table 1. The resource value is determined according to the contribution range into which the current contribution falls. If the resource value is a points multiplier, the redeemable rights can be determined according to the points multiplier.

[0095] Table 1

[0096] In steps S1101 to S1104 of this embodiment, the cloud receives the accuracy improvement, data volume, and training configuration parameters of the model uploaded by each vehicle terminal, and distributes resource values ​​accordingly to implement a resource incentive mechanism, thereby increasing the participation rate of vehicle terminals in model training and improving the accuracy of fault prediction.

[0097] In some instances, after the cloud completes the training of the current fault prediction model, it does not directly send the data packets of the current fault prediction model to each vehicle terminal, thus increasing the data transmission volume. Instead, the cloud first calculates the parameter difference between the current fault prediction model and the original fault prediction model, compresses the parameter difference to generate a binary differential packet, signs the differential packet, and then sends it to each vehicle terminal. Therefore, the vehicle terminal receives only 5-10% of the differential packet of the current fault prediction model, reducing the amount of data transmitted and improving the efficiency of model data delivery.

[0098] The technical solutions of the embodiments of this application, along with their corresponding effects and mechanisms of action, are described below. As shown in Table 2: Table 2

[0099] The following is a detailed description and explanation of the solutions in the embodiments of the present invention, using specific application examples: Taking the diagnosis of "engine vibration fault" as an example, the complete workflow of this invention is explained as follows: Scenario Description: A certain brand has 100,000 vehicles on sale and needs to establish an intelligent diagnostic model for engine vibration faults.

[0100] Phase 1: Initial Model Distribution The automaker used historical repair data to train a global fault prediction model: - Historical fault data: 5000 engine fault repair records - Global fault prediction model architecture: 1D-CNN-LSTM, 500K parameters - Initial accuracy: 75%; The cloud-based system will distribute the initialized original fault prediction model to the in-vehicle terminals of 100,000 vehicles via OTA: Model size of the original fault prediction model: 2MB (after INT8 quantization) - Distribution time: 7 days to complete the full push.

[0101] Phase 2: The vehicle-mounted terminal completes local training fault data collection and training. When a user encounters engine vibration while driving, data is collected based on trigger conditions: the vibration sensor detects an abnormal vibration pattern, or the user complains via voice that "the car is shaking." Data collection involves collecting sensor data (vibration, RPM, temperature, etc.) for 30 seconds before and after the vibration. This sensor data is encrypted and stored on the in-vehicle terminal. After accumulating 100 training fault data points, local training of the original fault prediction model is triggered. Local training consists of: 5 training rounds, batch size of 16, training learning rate of 0.001, and a training time of approximately 10 minutes (executed while charging).

[0102] Phase 3: In-vehicle terminal privacy protection and uploading: Privacy processing flow: 1. Clip the gradient of the updated fault prediction model: clip_norm=1.0; 2. Add differential privacy noise to the clipped gradient: epsilon=2.0, delta=1e-5; 3. Compress the noise-added gradient using Top-K compression: retaining the top 5% of gradient elements; 4. Anonymize user identity information in the compressed gradient: using a temporary ID to replace the vehicle VIN. Then, upload the privacy-processed gradient to the cloud, with the following upload conditions: first, a communication connection is detected; second, the target vehicle is charging and its battery level is greater than 50%. Uploaded gradient data size: compressed gradient: approximately 200KB and metadata (contribution metrics): approximately 1KB Phase 4: Cloud-based federated aggregation of multiple vehicle terminals: Aggregation cycle: Executed once every 24 hours. Vehicle terminal participation (for a given aggregation): Number of online vehicles: 8,523; Number of valid gradients: 6,847 (passed Byzantine test). Aggregation process: 1. Decrypt and verify the gradients uploaded by each vehicle terminal; 2. The Krum algorithm was used to detect outlier gradients, eliminating 127 suspicious gradients; 3. Aggregate gradients using FedAvg weighted averaging; 4. Update the parameters of the original fault prediction model using the aggregated gradient.

[0103] Furthermore, the update fault prediction model for each vehicle terminal was evaluated: the accuracy of the update fault prediction model was determined through the validation set: 75% → 82% (an improvement of 7%). Phase 5: Cloud-based model distribution and incentives for each in-vehicle terminal Model distribution: Generate an incremental update package (150KB) based on the current fault prediction model and the original fault prediction model, and conduct a canary release of the incremental update package: 1% → 10% → 100%. Full push is completed in 5 days.

[0104] Incentive distribution (example vehicle terminal P1): Vehicle terminal P1's contribution score is determined to be 85 points, vehicle terminal P1's points are determined to be 750 points, and vehicle terminal P1 can redeem benefits: 15% discount on the next maintenance.

[0105] In summary, the embodiments of this application, when applied to the "transmitter jitter fault" scenario, achieve the following results, as shown in Table 3: Table 3

[0106] As can be seen from the above results, the federated learning model training in this application embodiment can improve the fault prediction rate, rare fault identification rate, user participation rate, and number of participating vehicles, and can also increase the amount of data uploaded.

[0107] This application also provides a vehicle, which includes an in-vehicle terminal, a vehicle body, and at least one sensor. The in-vehicle terminal includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. The in-vehicle terminal can be any smart terminal, including a tablet computer, an in-vehicle computer, or similar device.

[0108] It is understood that the content of the above method embodiments is applicable to this vehicle embodiment. The specific functions implemented in this vehicle embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0109] This application also provides a vehicle fault prediction system, which includes at least one vehicle-mounted terminal and a cloud. The vehicle-mounted terminal is equipped with a first computer program, and the cloud is equipped with a second computer program. When the first computer program is executed by a processor, it implements the method executed by the vehicle-mounted terminal, and when the second computer program is executed by a processor, it implements the method executed by the cloud.

[0110] It is understood that the content of the above method embodiments is applicable to this vehicle fault prediction system. The specific functions implemented by this vehicle fault prediction system are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0111] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0112] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0115] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0116] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0117] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0118] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0120] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0121] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A vehicle fault prediction method, characterized in that, The method, executed by an in-vehicle terminal and connected to the cloud, includes the following steps: Acquire training fault data for the target vehicle; The preset original fault prediction model is trained based on the training fault data to obtain an updated fault prediction model and the original gradient data of the updated fault prediction model; wherein, the original fault prediction model is distributed and deployed in a lightweight manner by the cloud. The original gradient data is subjected to data perturbation processing to obtain encrypted gradient data; The encrypted gradient data is sent to the cloud, and the cloud is received as feedback of the model increment information for updating the fault prediction model based on the encrypted gradient data. The updated fault prediction model is updated based on the incremental information of the model to obtain the target fault prediction model; The target vehicle is used to predict faults using the target fault prediction model.

2. The method according to claim 1, characterized in that, The step of performing data perturbation processing on the original gradient data to obtain encrypted gradient data includes: The original gradient data is clipped according to preset clipping parameters to obtain the first gradient data; The first gradient data is subjected to noise processing to obtain the second gradient data; The second gradient data is compressed to obtain compressed gradient data; The compressed gradient data is anonymized to obtain the third gradient data; The third gradient data is encrypted to obtain the encrypted gradient data.

3. The method according to claim 1, characterized in that, The step of training a preset original fault prediction model based on the training fault data to obtain an updated fault prediction model and the original gradient data of the updated fault prediction model includes: Obtain the current operating status and current remaining energy of the target vehicle; The target vehicle is trained and evaluated based on the current operating status and the current remaining energy to obtain the training and evaluation results. If the training evaluation result indicates that the target vehicle meets the training conditions, the original fault prediction model is trained according to the training fault data and the preset training configuration parameters to obtain an updated fault prediction model. Obtain the original gradient data of the updated fault prediction model.

4. The method according to claim 3, characterized in that, After obtaining the original gradient data of the updated fault prediction model, the method further includes: Obtain the first prediction accuracy of the original fault prediction model and the second prediction accuracy of the updated fault prediction model; Obtain the amount of training fault data; The accuracy improvement value is determined based on the first prediction accuracy and the second prediction accuracy; The accuracy improvement value, the amount of data, and the training configuration parameters are sent to the cloud, and resource values ​​sent by the cloud based on the accuracy improvement value, the amount of data, and the training configuration parameters are received; wherein, the resource values ​​are used for resource exchange on the resource exchange platform of the target vehicle.

5. A vehicle fault prediction method, characterized in that, The method, executed by a cloud-based system and communicatively connected to at least one in-vehicle terminal, includes: The original fault prediction model is sent to the vehicle terminal; The system receives encrypted gradient data fed back by at least one vehicle-mounted terminal based on the original fault prediction model; wherein the encrypted gradient data is obtained by the vehicle-mounted terminal training a preset original fault prediction model based on collected training fault data to obtain original gradient data for updating the fault prediction model, and then performing data perturbation processing on the original gradient data. The original fault prediction model is optimized based on at least one of the encrypted gradient data to obtain the current fault prediction model. Determine model increment information based on the current fault prediction model and the original fault prediction model; The model increment information is sent to the vehicle terminal so that the vehicle terminal updates the updated fault prediction model according to the model increment information to obtain the target fault prediction model, and performs fault prediction on the target vehicle through the target fault prediction model.

6. The method according to claim 5, characterized in that, The step of optimizing the original fault prediction model based on at least one of the encrypted gradient data to obtain the current fault prediction model includes: Masking at least one of the encrypted gradient data yields at least one model gradient value. Anomaly assessment is performed on the gradient values ​​of the model to obtain anomaly assessment data; Based on the anomaly assessment data, reliable gradient values ​​are selected from the model gradient values; The original fault prediction model is optimized based on the reliable gradient value to obtain the current fault prediction model.

7. The method according to claim 6, characterized in that, The step of performing anomaly evaluation on the model gradient values ​​to obtain anomaly evaluation data includes: Calculate the distance between any two gradient values ​​of the model to obtain the gradient distance; Selected gradient values ​​are filtered from the model gradient values ​​based on a preset quantity and the gradient distance; The gradient distances are concatenated based on the selected gradient values ​​to obtain the anomaly assessment data.

8. The method according to claim 6, characterized in that, After sending the model incremental information to the vehicle terminal, the method further includes: Receive the accuracy improvement, data volume, and training configuration parameters sent by the vehicle terminal; Rarity assessment is performed on the updated fault prediction model to obtain rarity assessment data; The contribution of the vehicle terminal is evaluated based on the accuracy improvement, the amount of data, the training configuration parameters, and the rarity assessment data to obtain the current contribution. The vehicle terminal is allocated resources based on its current contribution level and a preset resource allocation table.

9. A vehicle, characterized in that, The vehicle includes: an on-board terminal, a body, and at least one sensor. The on-board terminal includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1 to 4.

10. A vehicle fault prediction system, the vehicle fault prediction system comprising at least one in-vehicle terminal and a cloud, characterized in that, The vehicle-mounted terminal is equipped with a first computer program, and the cloud is equipped with a second computer program. When the first computer program is executed by the processor, it implements the method of any one of claims 1 to 4, and when the second computer program is executed by the processor, it implements the method of any one of claims 5 to 8.

Images