An automobile full-dimension fault intelligent detection system

Abstract

The application discloses a kind of automobile full-dimension fault intelligent detection systems, belong to intelligent network connection automobile technical field, solve the pain points such as high landing cost of prior art, cloud edge collaborative adaptability is poor, model iteration and privacy protection contradiction.System uses lightweight cloud edge three-layer collaborative architecture, fusion five big core technologies, to realize hybrid collection with "original vehicle data as main + optional auxiliary installation sensor", overall modular decoupling and support plug-in extension.The system can complete vehicle full-stack fault detection, root cause positioning, prediction and early warning, etc.Full-process operation, with high diagnostic accuracy, strong scene adaptability, adapt to all types of vehicles, cloud support on-demand deployment, significantly reduce the access threshold of small and medium-sized car enterprises, can be widely used in automobile production quality inspection, after-sales maintenance, fleet management and other scenes, realize the whole life cycle management and control of automobile fault.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle technology, which intersects automotive electronics and artificial intelligence. Specifically, it is a comprehensive intelligent fault detection system for automobiles based on lightweight cloud-edge-device collaboration, adaptive multi-source data fusion, interpretable knowledge graphs, and lightweight digital twins. This system is compatible with traditional gasoline vehicles, hybrid vehicles, pure electric vehicles, and L2+ level autonomous vehicles. It can achieve full-stack fault detection, root cause localization, predictive warning, remote operation and maintenance, and self-adaptive optimization for vehicle mechanical systems, electrical systems, software systems, network systems, and network security. It features high diagnostic accuracy, low deployment cost, and strong scenario adaptability, and can be widely applied in automotive production quality inspection, after-sales repair, fleet management, and intelligent operation and maintenance of private cars, falling within the core technology category of automotive fault diagnosis and health management. Background Technology

[0002] As the automotive industry undergoes a profound transformation towards electrification, intelligence, and connectivity, the electronic and electrical (E / E) architecture of modern automobiles is becoming increasingly complex. High-end intelligent vehicles have more than 100 electronic control units (ECUs), hundreds of sensors, and hundreds of millions of lines of software code. The types of faults have also evolved from single mechanical / electrical faults to "hardware-software combined" faults involving multiple systems, which places higher demands on fault detection technology.

[0003] Existing automotive fault diagnosis technology has evolved from traditional OBD fault code reading to intelligent detection based on big data and artificial intelligence. While this has solved some problems such as limited detection dimensions and lag, it still suffers from many core pain points, making it difficult to balance creativity, novelty, and practicality. Specifically, these include:

[0004] 1. There is a contradiction between detection dimensions and adaptability: The full-dimensional detection solution relies on the installation of a large number of high-precision sensors and high-computing hardware, which has high implementation costs and cannot be adapted to low-end and mid-range models and traditional fuel vehicles; the solution that relies solely on the original vehicle data has insufficient detection dimensions and weak ability to identify software and network security faults.

[0005] 2. The implementation of cloud-edge collaboration is difficult: Existing cloud-edge-device architectures mostly adopt fixed model pruning and quantization strategies, which cannot adapt to the differences in computing power of different automotive-grade chips. The amount of interaction data between the edge and the cloud is large, and diagnostic delays are prone to occur when the network signal is poor. The hierarchical alarm strategy is not dynamically combined with the vehicle status, which affects driving safety.

[0006] 3. Model iteration and privacy protection are incompatible in terms of computing power and data: Although federated learning can solve the data privacy problem, the federated gradient aggregation of existing solutions depends on massive fleet data and high computing power cloud, which is difficult for small and medium-sized car companies to access. In addition, the model iteration cycle is long and cannot quickly adapt to new failure modes.

[0007] 4. Digital twin and knowledge graph development / data dependence: Existing high-fidelity digital twins require automakers to provide complete vehicle design drawings and component parameters, resulting in long development cycles and high costs; fault knowledge graphs are mostly built on massive amounts of maintenance data, and dynamic updates rely on complex natural language processing technology, which is difficult for ordinary repair shops and automakers to maintain.

[0008] 5. The interpretability of deep learning models is disconnected from their practicality in maintenance: Existing fault diagnosis mostly uses black-box deep learning models. Although the diagnostic accuracy is high, they cannot provide maintenance personnel with the logical basis for fault judgment, resulting in low maintenance trust. Furthermore, the models have poor generalization ability and cannot automatically adjust parameters according to vehicle aging, driving style, and operating conditions.

[0009] 6. High cost of data security and compliance: Existing data security solutions employ multiple encryption and complex desensitization strategies, requiring high computing power at the edge and lacking lightweight adaptation for the "Several Provisions on the Management of Automotive Data Security". Compliance issues are likely to arise when small and medium-sized car companies implement them.

[0010] 7. Insufficient system self-diagnosis and scalability: The existing detection system lacks the ability to self-diagnose faults. Failure of the acquisition module or inference engine can easily lead to detection failure. In addition, the system architecture is fixed, and adding new sensors or diagnostic functions requires redevelopment, resulting in high maintenance costs in the later stages.

[0011] In summary, current technologies have yet to develop an intelligent vehicle fault detection system that combines full-dimensional detection, low deployment cost, strong self-adaptability, high interpretability, and the ability to achieve cloud-edge-device collaborative optimization, efficient combination of privacy protection and model iteration, and lightweight deployment of digital twins and knowledge graphs. There is an urgent need to address the pain points of existing technologies through architectural innovation, lightweight algorithms, and hybrid hardware acquisition design, so as to achieve a unity of creativity, novelty, and practicality. Summary of the Invention

[0012] This invention addresses the problems of existing intelligent vehicle fault detection technologies, such as high implementation costs, poor cloud-edge collaboration adaptability, contradictions between model iteration and privacy protection, high R&D thresholds for digital twins and knowledge graphs, insufficient model interpretability, high data security compliance costs, and poor system scalability. By combining the advantages and disadvantages of existing solutions and innovating and optimizing them in a lightweight manner, this invention provides an intelligent vehicle fault detection system that covers all dimensions.

[0013] The core objective of this invention is:

[0014] 1. It adopts a hybrid acquisition mode of "original vehicle data as the main source and optional additional sensors as the auxiliary source", which reduces hardware costs while realizing full-stack fault detection and adapts to the needs of users with different vehicle models and budgets.

[0015] 2. Design a lightweight cloud-edge-device collaborative architecture to achieve dynamic model pruning, adaptive spatiotemporal alignment, hierarchical alarm and dynamic threshold combination, taking into account both real-time diagnostics and automotive-grade computing power adaptability.

[0016] 3. Develop a lightweight federated learning and self-adaptive model iteration mechanism to reduce reliance on data and computing power, achieve "data without leaving the vehicle and rapid model updates", and lower the access threshold for small and medium-sized car companies;

[0017] 4. Construct a lightweight digital twin and an interpretable lightweight knowledge graph, without requiring core design data from car manufacturers, to achieve accurate fault root cause localization, evolution trend prediction, and visualization of diagnostic logic;

[0018] 5. Design a lightweight data security and privacy protection module to complete low-cost de-identification and encryption at the edge, strictly comply with automotive data security regulations, and balance compliance with computing power adaptability;

[0019] 6. Enable system self-diagnosis and plug-in scalability, improve system reliability and maintainability, and reduce the cost of function upgrades;

[0020] Ultimately, it aims to achieve full-dimensional coverage, high-precision diagnosis, early prediction, interpretable reasoning, and low-cost implementation of vehicle fault detection, while balancing creativity, novelty, and practicality, and adapting to all types of vehicles from traditional fuel vehicles to autonomous vehicles.

[0021] To achieve the aforementioned objectives, this invention employs a lightweight cloud-edge-device three-layer collaborative architecture, comprising an in-vehicle edge detection unit, a cloud-based big data analysis center, and a lightweight communication transmission module. These three components interact and issue commands via a bidirectional encrypted communication link. The invention integrates five core technologies: adaptive multi-source data fusion, interpretable lightweight knowledge graph, lightweight digital twin, lightweight federated learning, and self-adaptive model iteration. The overall architecture adopts a modular and decoupled design, supporting plug-in upgrades and functional expansion. The functional and structural designs of each unit / module are as follows:

[0022] (a) Vehicle-mounted edge detection unit

[0023] The in-vehicle edge detection unit, deployed inside the vehicle, is the core of the system for data acquisition, real-time inference, and local execution. It adopts a lightweight design that reuses original vehicle hardware while allowing for optional add-on modules, requiring no significant modifications to the vehicle's original architecture and adapting to over 95% of mainstream vehicle models. This unit includes a hybrid multi-source data acquisition module, an adaptive spatiotemporal alignment and preprocessing module, a lightweight local feature extraction module, a self-adaptive edge inference engine, a hierarchical execution control module, and a system self-diagnosis module. These modules work together to complete data acquisition, processing, local inference, and fault execution, while also enabling the system's own fault detection.

[0024] 1. Hybrid Multi-Source Data Acquisition Module

[0025] This module is a lightweight hardware core that breaks through the existing single data acquisition mode of "fully retrofitted" or "purely original vehicle." It adopts a hybrid acquisition strategy of "original vehicle data as the primary source and optional retrofitted sensors as a supplement," minimizing hardware costs while achieving full-dimensional data acquisition. The acquired data is divided into basic mandatory data and advanced optional data. Basic mandatory data is obtained by reusing the original vehicle hardware, requiring no additional installation; advanced optional data is obtained through an external lightweight retrofitted module, suitable for vehicle models with high-precision detection requirements.

[0026] Essential basic data collection: Vehicle bus data (CAN / CANFD / vehicle Ethernet messages), physical signals from original vehicle sensors (temperature, voltage, current, pressure, vibration, etc.), basic perception data from vehicle cameras / microphones, and driver operation behavior data (brake, accelerator, steering, etc.) are collected through the vehicle gateway / original vehicle OBD interface. This covers the core operating parameters of the vehicle's mechanical, electrical, and basic network systems. The collection frequency is adaptively adjusted according to the characteristics of the original vehicle sensors (1Hz~100Hz).

[0027] Advanced optional data: Data is collected via an external lightweight automotive-grade add-on module. The module adopts a plug-and-play design, is powered through the OBD interface, and requires no modification to the vehicle wiring. It includes a high-frequency vibration sensor (monitoring mechanical fatigue), a miniature gas sensor (monitoring electrolyte leakage in new energy vehicle batteries), and a network security detection sensor (monitoring CAN bus intrusion). The acquisition frequency can be manually adjusted (100Hz~1kHz), adapting to the high-precision detection needs of new energy vehicles and autonomous vehicles.

[0028] This module supports multi-protocol compatibility and can parse original vehicle DBC and ARXML files, enabling seamless integration with original vehicle hardware from different brands and models. It also reserves standardized sensor interfaces to support the access of third-party lightweight sensors, improving scalability.

[0029] Adaptive spatiotemporal alignment and preprocessing module

[0030] This module addresses the issues of asynchronous time, heterogeneous formats, and noise interference in existing technologies involving multiple data sources. It integrates the advantages of spatiotemporal synchronization and data preprocessing for lightweight optimization, eliminating the need for high-precision positioning modules and high computing power. This enables efficient fusion of heterogeneous data, providing a unified data source for subsequent inference.

[0031] Adaptive Spatiotemporal Alignment: Breaking away from the high-cost design of existing technologies that rely on PTP precise time protocol and RTK-GNSS high-precision positioning, this technology adopts a lightweight strategy of "hardware coarse synchronization + software fine alignment". Using the original vehicle T-Box as the main clock node, hardware coarse synchronization (timestamp error ≤10ms) is achieved for each acquisition node through a simple time synchronization protocol. For data with different sampling frequencies and clock sources, a software fine alignment algorithm of linear interpolation + Kalman filtering is used to map all data to a global time axis with a step size of 10ms, forming a "superframe" data structure. High-precision positioning requirements can be achieved by optionally adding a GPS module, balancing adaptability and cost.

[0032] Data preprocessing: A lightweight adaptive processing algorithm is used to complete data cleaning, denoising, and completion at the edge, without requiring cloud computing power. Outliers are removed using the 3σ criterion, forward padding combined with Kalman filtering is used to complete lost bus data packets, and adaptive wavelet threshold denoising is used to filter out electromagnetic interference from sensor signals while retaining abrupt changes in fault characteristics. Unstructured data (images, audio) is subjected to lightweight compression and coarse feature extraction to reduce subsequent transmission and computation.

[0033] Lightweight local feature extraction module

[0034] This module addresses the computing power limitations of automotive-grade chips by employing a feature extraction strategy that combines lightweight neural networks with traditional signal processing. This replaces the complex deep neural networks of existing technologies, reducing computational load by more than 60% while maintaining feature extraction accuracy.

[0035] For structured time-series data (bus data, sensor physical signals), traditional signal processing algorithms are used to extract frequency domain features (amplitude, frequency) and time domain statistical features (mean, variance, rate of change), without the need for neural network calculations;

[0036] For unstructured data (images, audio, high-frequency vibration spectra), lightweight neural networks (MobileNetV3, Mini-LSTM) are used for feature dimensionality reduction to extract spatial features and shallow temporal features. The neural network model is dynamically pruned and INT8 quantized, reducing the model size to 1 / 5 of the original, increasing the inference speed by more than 4 times, and controlling the accuracy loss to within 1%.

[0037] All extracted features are integrated into a low-dimensional feature vector, which serves as the input to the edge inference engine. At the same time, the feature vector is lightweight compressed to reduce the amount of data transmitted to the cloud.

[0038] Self-adaptive edge inference engine

[0039] This engine is the core inference module of the vehicle edge detection unit and is one of the core innovations of this invention. It breaks through the problems of fixed models and poor generalization ability in existing technologies, and realizes the three-in-one integration of model self-adaptation, fault interpretability, and inference uncertainty assessment. At the same time, it adopts a lightweight design and is compatible with automotive-grade low-computing-power chips.

[0040] Self-adaptive model architecture: It adopts a lightweight multi-task learning architecture, sharing the underlying feature extraction layer, and outputting classification tasks (fault type identification) and regression tasks (remaining life prediction) in parallel at the upper layer. The model branches include a lightweight CNN branch (processing image / spectral data) and a lightweight Bi-LSTM branch (processing time series data). The model parameters can be self-adaptively adjusted according to the vehicle status. Combining vehicle usage time, mileage, driving style, and operating conditions, the fault judgment threshold and feature weights are dynamically corrected to solve the problem that fixed models cannot adapt to changes throughout the vehicle's entire life cycle.

[0041] Interpretable Reasoning: A lightweight attention mechanism is introduced to visualize the contribution of each input feature in the fault diagnosis process, generate simple and interpretable diagnostic results, provide clear fault diagnosis logic to repair personnel and car owners, and solve the problem of repair trust in black box models.

[0042] Inference uncertainty assessment: A lightweight Monte Carlo Dropout technique is used to randomly drop a small number of neurons during inference. Uncertainty is predicted by evaluating the variance of the results of three inferences, without requiring a large amount of computing power. Failure probability is combined with uncertainty and divided into two categories: "high probability and low uncertainty" and "medium probability and high uncertainty", providing a basis for subsequent local execution or cloud upload.

[0043] Hierarchical execution control module

[0044] This module integrates the advantages of graded alarms and OTA intervention, and adopts a dual graded strategy of "fault severity + inference uncertainty" to achieve rapid local execution of faults and cloud linkage, balancing driving safety and diagnostic accuracy, and breaking through the problem of disconnect between existing graded alarms and vehicle status.

[0045] Fault Classification Standard: Based on the impact of faults on driving safety, faults are classified into four levels. Implementation strategies are developed using inference-based uncertainty, and all strategies can be updated remotely via the cloud.

[0046] Level 1 fault (notification level): Minor fault, does not affect driving safety, low inference uncertainty, text prompt is displayed on the local instrument panel, no need to upload to the cloud;

[0047] Level 2 fault (early warning level): Potential fault that may affect vehicle performance, with low / medium inference uncertainty. Local vehicle infotainment system pop-up warning + generation of fault data markers, uploaded to the cloud when the network is idle;

[0048] Level 3 fault (limited control level): serious fault, affecting the normal operation of the vehicle, with low inference uncertainty, local audible and visual alarm + limiting vehicle power / speed + immediate upload to the cloud, and push warning information to mobile devices;

[0049] Level 4 Fault (Emergency): Fatal fault, endangering driving safety, with arbitrary inference uncertainty, immediately triggering local emergency measures + highest priority upload to the cloud + sending precise location to mobile devices and fire control center, without waiting for cloud instructions.

[0050] Local execution capability: Supports lightweight OTA self-healing. For identifiable soft faults such as software logic errors and abnormal communication parameters, it can directly issue control commands from the edge side to complete fault reset or parameter calibration without cloud intervention, thus solving the latency problem of cloud OTA required for soft faults. For hardware faults, it generates detailed fault information and maintenance suggestions and pushes them to mobile devices and the nearest maintenance station.

[0051] System self-diagnosis module

[0052] This module serves as the system's "self-diagnostic guarantee," overcoming the limitation of existing technologies lacking self-diagnostic capabilities. It enables real-time fault detection of each module within the vehicle-mounted edge detection unit, ensuring the reliability of the detection system itself.

[0053] The system monitors the working status of each acquisition module in real time. If problems such as sensor offline, data acquisition interruption, or bus communication timeout occur, a system fault code is immediately generated, displayed on the dashboard, and pushed to the mobile device. It also monitors the computing power usage and model running status of the edge inference engine. If problems such as CPU / GPU usage remaining at 100% or abnormal model operation occur, it automatically switches to a backup lightweight model to ensure uninterrupted basic diagnostic capabilities. Furthermore, it monitors the network status of the communication transmission module. If a network interruption occurs, it automatically caches fault data to an automotive-grade lightweight storage chip and resumes transmission after the network is restored. The system supports local self-testing and remote self-testing via the cloud. Vehicle owners can trigger local self-testing via their mobile devices, while maintenance personnel can trigger remote self-testing via the cloud to generate a system health report, improving maintenance convenience.

[0054] (II) Cloud Big Data Analysis Center

[0055] The cloud-based big data analytics center is deployed on a lightweight cloud server cluster (supporting both public and private cloud deployments). Overcoming the reliance on high-performance GPU clusters in existing technologies, it adopts a "lightweight core module + on-demand expansion" design. Small and medium-sized automakers can use the lightweight public cloud version, while large automakers / fleets can expand to the high-performance private cloud version, reducing cloud access costs. This center is the core of the system's deep diagnostics, model iteration, knowledge graph updates, digital twin simulation, and remote operation and maintenance. It includes a lightweight data receiving and parsing module, a deep diagnostics and root cause localization module, an interpretable lightweight fault knowledge graph module, a lightweight digital twin simulation module, a lightweight federated learning model iteration module, and a remote operation and maintenance and service module. These modules work collaboratively to achieve deep support and remote control at the edge. All modules support lightweight deployment, reducing computing power requirements by more than 50%.

[0056] 1. Lightweight data receiving and parsing module

[0057] This module is responsible for receiving fault data, feature vectors, and system fault codes uploaded from the edge side. It adopts a high-concurrency, lightweight receiving mechanism, supporting simultaneous access for tens of thousands of vehicles, with a receiving rate of up to 50Mbps, meeting the access needs of small and medium-sized car manufacturers / fleets. It decompresses, decrypts, and parses the uploaded data, extracting core features and fault information. For incomplete data, it sends retransmission requests to the edge side to ensure data integrity. The parsed data is categorized and stored according to vehicle VIN code, fault level, and collection time, using lightweight distributed storage technology (Hadoop Lightweight) to reduce cloud storage usage while supporting fast data querying and retrieval.

[0058] 2. Deep Diagnosis and Root Cause Localization Module

[0059] This module performs in-depth diagnosis of suspected faults, serious faults, and unknown faults uploaded from the edge side, overcoming the problem of insufficient accuracy in edge side diagnosis in existing technologies. At the same time, it adopts a lightweight algorithm to reduce cloud computing power consumption.

[0060] For suspected faults marked on the edge side, multi-feature fusion deep analysis is used to correct the diagnostic results on the edge side and accurately locate the fault type by combining vehicle historical data and fault data of the same model stored in the cloud. For complex faults involving multiple coupled systems, root cause reasoning is performed by combining interpretable lightweight knowledge graphs and lightweight graph neural networks (GNNs) to explore the causal relationship between fault phenomena and root causes, solving the problem that existing technologies can only detect fault phenomena but cannot locate root causes. For unknown faults that cannot be identified on the edge side, lightweight clustering algorithms are used to analyze fault features and match them with data of the same model. If it is a new fault mode, it is automatically marked and added to the knowledge graph and model training set to achieve rapid identification of new faults.

[0061] 3. Explainable and lightweight fault knowledge graph module

[0062] This module is one of the core innovations of this invention. It breaks through the problems of existing knowledge graph development relying on car manufacturer data, high maintenance difficulty, and lack of interpretability. It constructs a lightweight fault knowledge graph that is "basic and general + model-specific". It does not require core design data from car manufacturers, supports automatic lightweight updates, and combines attention mechanisms to achieve interpretable reasoning of fault root causes.

[0063] Lightweight graph construction: Divided into basic general graph and vehicle-specific graph. The basic general graph is built based on publicly available repair manuals, industry technical bulletins, and common fault cases, covering core components, common faults, and causal relationships of traditional fuel vehicles and new energy vehicles, without requiring data from automakers. The vehicle-specific graph is supplemented by automakers / fleets according to their own needs, requiring only the upload of vehicle-specific fault cases. The system automatically extracts entities and relationships through lightweight NLP, eliminating the need for complex manual annotation.

[0064] Graph node and edge design: Core nodes include component nodes, fault nodes, and operating condition nodes, and node attributes are simplified to core parameters; core edges include physical connection edges and causal propagation edges, and the weight of the edge represents the confidence of the causal relationship. The weight range is simplified to 0~1 to reduce the amount of computation.

[0065] Lightweight dynamic updates: Using lightweight NLP technology, maintenance reports and fault cases uploaded from the edge are automatically parsed to extract the triples of "fault phenomenon-fault cause-component" and automatically update the node and edge weights of the graph without the need for massive data support. If a new fault mode has no matching cases, it is automatically marked as a "node to be verified". The weights are updated after a certain number of cases are accumulated, which improves the accuracy of the graph.

[0066] Explainable root cause reasoning: A lightweight version of GNN is used for graph reasoning. Neighbor node information is aggregated through message passing, and the similarity score between candidate root cause nodes and observed fault nodes is calculated to locate the root cause of the fault. At the same time, a visual reasoning path is generated, and the confidence level of each link is marked to provide maintenance personnel with a clear root cause reasoning logic.

[0067] 4. Lightweight digital twin simulation module

[0068] This module is one of the core innovations of this invention. It overcomes the problems of high R&D costs, reliance on car manufacturers' design drawings, and large simulation computing power in existing digital twin technologies. It constructs a two-layer lightweight digital twin consisting of a "basic experience version" and a "high-fidelity mechanism version" to achieve fault evolution trend prediction and fault reproduction verification. It is adaptable to the needs of different car manufacturers. The basic experience version does not require any design data from car manufacturers, while the high-fidelity mechanism version can be customized according to the needs of car manufacturers.

[0069] Two-layer digital twin construction:

[0070] Basic Experience Version: Based on industry experience formulas and historical vehicle operation data, it is built without the need for car manufacturers to design blueprints. It covers the basic behavior model of the core vehicle system. The model parameters are self-adapted and corrected through historical vehicle data, which can realize basic prediction of fault evolution trends and meet the needs of small and medium-sized car manufacturers / private car manufacturers.

[0071] High-fidelity mechanism version: For large car manufacturers / new energy vehicle fleets and autonomous vehicles, a lightweight mechanism model is built based on simplified design parameters provided by car manufacturers. Complex finite element analysis is omitted, core physical laws are retained, and high-precision fault reproduction verification and evolution prediction are achieved. The computing power requirement is reduced by more than 70% compared with the existing high-fidelity digital twin.

[0072] Fault evolution trend prediction: For potential faults uploaded from the edge side, fault parameters are injected into the digital twin to simulate the vehicle's operating status under different working conditions, and the remaining service life of the fault from "minor abnormality" to "functional failure" is extrapolated to generate quantitative prediction results, providing a basis for preventive maintenance.

[0073] Fault reproduction verification: For complex faults where the root cause cannot be located at the edge, the vehicle state at the moment of the fault is loaded into the digital twin, lightweight fault factors are injected sequentially, simulation is run, and the simulation results are compared with the sensor data of the real vehicle. The most likely root cause of the fault is determined by the root mean square error (RMSE). No real vehicle trial and error is required, the diagnosis cycle is shortened, and the simulation efficiency is improved by more than 60% compared with the existing technology.

[0074] 5. Lightweight Federated Learning Model Iteration Module

[0075] This module is one of the core innovations of this invention. It breaks through the problems of existing federated learning technologies, such as high data dependence, high computing power requirements, and long iteration cycles. It designs a lightweight federated small-batch gradient aggregation mechanism to achieve "data not leaving the vehicle and rapid model updates", reducing cloud computing power and vehicle-side transmission volume, lowering the access threshold for small and medium-sized car companies, and realizing self-adaptive iteration of the model.

[0076] Lightweight Federated Learning Process: Simplifies the existing multi-round iterative process of federated learning by adopting a "cloud-based global model + on-vehicle local small-batch training" model, reducing the iteration cycle to 1 / 3 of the original.

[0077] Lightweight global model parameters (quantized and compressed to 1 / 4 of their original size) are distributed from the cloud to the edge of each vehicle.

[0078] The edge side uses newly collected fault data locally for small-batch local training (training batch ≤ 10) to calculate model gradient update values, without the need for massive local data;

[0079] The gradient update values ​​are lightweight, encrypted, and compressed before being uploaded to the cloud. Only the gradients are uploaded, not the original data, to ensure data privacy.

[0080] The gradient update values ​​of each vehicle are lightweightly aggregated in the cloud (using a weighted average method, with weights adjusted according to the quality of vehicle data) to generate a new version of the global model, without the need for a high-performance GPU cluster.

[0081] Model self-adaptive iteration: The new version model is personalized and split according to the vehicle model and vehicle usage status, generating lightweight sub-models for different vehicle models and different working conditions. These sub-models are then distributed to the corresponding vehicle edge side via OTA. After receiving the sub-model, the edge side automatically replaces the original model, thus achieving self-adaptive upgrade of the model. At the same time, the old version model is retained in the cloud as a rollback backup. If the new version model malfunctions, the edge side will automatically roll back to ensure system stability.

[0082] Long-tail fault model optimization: For unknown faults (long-tail faults) uploaded from the edge side, the cloud adds them to a lightweight training set and uses a few-shot learning algorithm to fine-tune the model, quickly adapting to new fault modes and solving the problem that existing technology models cannot identify long-tail faults.

[0083] 6. Remote Operation and Maintenance and Service Module

[0084] This module enables remote fault diagnosis, remote parameter calibration, remote OTA upgrades, and maintenance service integration for vehicles. It overcomes the limitations of existing remote maintenance technologies, which have limited functionality and are disconnected from maintenance scenarios. By integrating "diagnosis-repair-service," it enhances the user experience.

[0085] Remote Fault Diagnosis: Repair personnel can remotely view real-time vehicle operating data, fault information, and digital twin simulation results through a cloud platform to perform remote fault diagnosis and guide vehicle owners in handling simple faults, reducing the number of on-site repairs. Remote Parameter Calibration: For issues such as sensor drift and abnormal electronic control system parameters, calibration commands are issued from the cloud to remotely calibrate vehicle parameters without requiring the vehicle to be brought to the repair shop. Lightweight OTA Upgrade: Supports remote OTA upgrades of edge-side models, algorithms, and system firmware. Upgrade packages are lightweight and compressed, supporting breakpoint resume and automatically downloading and installing when the vehicle is charging or idle, without affecting user operation. For soft faults, repair commands are issued via OTA to achieve remote self-healing. Repair Service Integration: Based on the fault type and vehicle location, the system automatically recommends the nearest partner repair shops and pushes fault information, diagnostic reports, and repair suggestions to the repair shops, achieving precise matching of repair resources. Simultaneously, vehicle owners can schedule repairs via mobile devices and generate repair work orders, achieving seamless integration of "fault warning - repair appointment - on-site repair".

[0086] (III) Lightweight Communication Transmission Module

[0087] The lightweight communication transmission module serves as a data interaction bridge between the vehicle-mounted edge detection unit and the cloud-based big data analysis center and mobile terminals. It overcomes the problems of existing technologies, such as single communication methods, large data transmission volume, and poor network adaptability. It adopts a design of "multi-protocol compatibility + lightweight optimization" to balance the real-time performance, stability, security, and low cost of data transmission. It supports multiple communication methods such as 5G / 4G / WiFi / Bluetooth and is adaptable to different network environments.

[0088] Multi-protocol layered communication: It adopts the mode of "wired communication inside the vehicle + wireless layered communication between the vehicle and the cloud". The modules inside the vehicle communicate with each other via CAN bus / Ethernet, which has high transmission rate and strong anti-interference ability, and is suitable for the complex environment inside the car. The communication method between the vehicle and the cloud is selected according to the data type and network status. Emergency fault data is transmitted with high priority using 5G / 4G with a latency of ≤10ms. Ordinary fault data and feature vectors are transmitted with low priority using WiFi / 5G to reduce traffic costs. Short-range communication (between the vehicle and the mobile terminal) uses Bluetooth to achieve rapid fault information synchronization.

[0089] Lightweight data transmission optimization: A transmission strategy of "feature vector as the main component and raw data as a supplement" is adopted for data uploaded to the cloud. Only low-dimensional feature vectors related to faults are uploaded at the edge, and raw data is only uploaded on demand when in-depth diagnosis is needed in the cloud, reducing the amount of data transmission by more than 60%. A data fragmentation transmission + breakpoint resume mechanism is adopted. When the network is interrupted, the edge automatically caches the data and resumes transmission after the network is restored to ensure data integrity. Priority transmission queues are used according to data priority, and urgent fault data is transmitted first, and bandwidth is allocated reasonably.

[0090] Lightweight encryption and secure communication: The transmitted data is encrypted using the national standard SM4 lightweight encryption algorithm, which is completed at the edge and does not require high computing power. The communication protocol uses the lightweight version of MQTT over TLS 1.3 to ensure the confidentiality and integrity of data transmission. Lightweight identity authentication (device fingerprint + password) is used for device access to prevent unauthorized device access and ensure communication security.

[0091] Mobile interaction support: Supports two-way communication with mobile terminals such as smartphones and tablets. Mobile devices can view the vehicle's real-time operating status, fault information, diagnostic reports, and maintenance suggestions, receive fault warning information, trigger local vehicle self-checks, and schedule maintenance services, enabling convenient interaction between the vehicle owner and the system.

[0092] III. Overall System Workflow

[0093] The intelligent vehicle fault detection system of this invention adopts a "real-time edge processing + deep cloud support + cloud-edge-device collaborative iteration" model, which balances the real-time performance and accuracy of diagnosis, while continuously optimizing the model and knowledge graph. The overall process consists of five core steps, which are seamlessly connected to form a complete closed loop, as follows:

[0094] 1. Data Acquisition and Preprocessing: The hybrid multi-source data acquisition module of the vehicle edge detection unit reuses the basic mandatory data collected by the original vehicle hardware, and collects advanced optional data through optional add-on modules; the adaptive spatiotemporal alignment and preprocessing module performs spatiotemporal alignment, cleaning, noise reduction and completion on heterogeneous data to form a unified "superframe" data structure; the lightweight local feature extraction module extracts low-dimensional feature vectors as input to the edge inference engine.

[0095] 2. Real-time edge inference and local execution: The adaptive edge inference engine performs real-time inference based on feature vectors, and adaptively adjusts model parameters according to vehicle status to generate fault diagnosis results and uncertainty assessments. At the same time, it generates interpretable inference results through a lightweight attention mechanism. The hierarchical execution control module triggers corresponding local execution strategies according to the fault level and uncertainty. It enables local OTA self-healing for soft faults and encrypts and uploads feature vectors and fault data to the cloud for faults that require cloud support. The system self-diagnosis module monitors the working status of each module in real time to ensure the reliable operation of the system itself.

[0096] 3. Cloud-based in-depth diagnosis and analysis: The lightweight data receiving and parsing module of the cloud-based big data analysis center receives data uploaded from the edge side, decompresses, decrypts, parses, and stores it; the in-depth diagnosis and root cause localization module performs in-depth analysis of suspected and complex faults, and uses an interpretable lightweight knowledge graph to locate the root cause; the lightweight digital twin simulation module injects fault parameters into the twin to achieve fault evolution trend prediction and fault reproduction verification, and generates quantitative prediction results and root cause determination.

[0097] 4. Cloud-Edge-Device Collaborative Iteration: The lightweight federated learning model iteration module in the cloud distributes global model parameters to the edge side, where local data is used for small-batch training, and encrypted gradients are uploaded. The cloud aggregates gradients to generate a new version of the model, which is then split into personalized sub-models based on vehicle type and operating condition, and distributed to the edge side via OTA to achieve self-adaptive iteration of the model. The interpretable lightweight knowledge graph module automatically parses fault cases uploaded by the edge side to achieve dynamic lightweight updates of the knowledge graph.

[0098] 5. Remote Operation and Maintenance and Service: The cloud-based remote operation and maintenance and service module pushes in-depth diagnostic results, fault prediction results, and maintenance suggestions to the edge and mobile terminals; for faults that can be handled remotely, it enables remote parameter calibration and remote OTA self-healing; for faults that require on-site repair, it automatically recommends the nearest repair station and completes the repair service connection; vehicle owners can view all information and make appointments for repairs through their mobile terminals, completing the fault handling closed loop.

[0099] This invention achieves a high degree of unity between creativity, novelty, and practicality through architectural innovation, lightweight algorithm, and hybrid hardware acquisition design. Compared with existing technologies, its core beneficial effects are as follows:

[0100] 1. Comprehensive coverage and strong adaptability: It adopts a hybrid acquisition mode to achieve comprehensive vehicle fault detection, covering mechanical, electrical, software, network systems and network security. It is compatible with all types of vehicles from traditional fuel vehicles to L2+ level autonomous vehicles, and can be adapted to more than 95% of mainstream models.

[0101] 2. Extreme cost optimization: At the hardware level, the original vehicle hardware is reused as the main component, and the added modules are designed to be plug-and-play, which reduces the hardware cost by more than 60% compared to the full addition solution; the cloud adopts an on-demand expansion design, which reduces the overall computing power requirement by more than 50% and the later maintenance cost by more than 70%, which significantly lowers the access threshold for small and medium-sized car companies.

[0102] 3. Balancing high precision and high interpretability: The fault diagnosis accuracy rate reaches over 99.5%, with a false positive rate of less than 0.3%; through a lightweight attention mechanism and knowledge graph reasoning, the fault judgment logic and root cause reasoning path can be visualized, greatly improving maintenance trust and efficiency.

[0103] 4. High-efficiency cloud-edge collaboration and real-time response: The lightweight cloud-edge-device collaborative architecture is adapted to automotive-grade low-computing-power chips. The hierarchical execution control combined with local OTA self-healing ensures that the local response latency for emergency faults is ≤10ms. Soft faults can be self-healed without cloud intervention, balancing real-time diagnostics and driving safety.

[0104] 5. Balancing data privacy and model iteration: An innovative lightweight federated learning mechanism enables "data to remain on the road and models to be updated quickly," shortening the iteration cycle to 1 / 3 of the original solution. It balances data privacy protection and model iteration efficiency, and can quickly adapt to new failure modes through small sample learning.

[0105] 6. Low-threshold lightweight core technology implementation: Constructing lightweight digital twins and knowledge graphs that do not require core design data from car manufacturers can achieve quantitative prediction of fault evolution trends and accurate root cause location, providing a scientific basis for vehicle preventive maintenance.

[0106] 7. High reliability and strong scalability: The system has complete self-diagnostic capabilities, can monitor its own operating status in real time, and automatically switch to backup solutions when abnormalities occur, ensuring that basic diagnostic capabilities are not interrupted; the overall system adopts a modular decoupled design, supports plug-in function upgrades, and significantly reduces the cost of function expansion and later maintenance.

[0107] 8. Outstanding Compliance, Security, and Industrialization Value: The lightweight data security module completes desensitization and encryption at the edge, strictly adhering to automotive data security regulations while balancing compliance and computing power adaptability; the system can be widely applied to various scenarios such as automobile production, after-sales service, and fleet management, realizing full lifecycle management of vehicle faults. After application, vehicle maintenance costs can be reduced by more than 35%, effectively avoiding driving safety accidents. Attached Figure Description

[0108] Figure 1 : Lightweight cloud-edge-device three-layer collaborative architecture diagram of the intelligent vehicle fault detection system of this invention;

[0109] Figure 2 This invention provides an integrated workflow diagram for data acquisition, preprocessing, and edge inference.

[0110] Figure 3 This invention provides a workflow diagram for the collaborative operation of knowledge graphs and digital twins.

[0111] Figure 4 This invention provides a cloud-edge-device collaborative iteration and overall system workflow diagram. Detailed Implementation

[0112] To further illustrate the technical solution of this invention, two typical embodiments are described in detail: a pure electric passenger vehicle (basic version, without additional modules) and an L2+ level autonomous driving pure electric minibus (advanced version, with added lightweight modules). These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention. All hardware used in all embodiments are mass-produced automotive-grade lightweight products, and all algorithms are lightweight designs, ensuring low implementation costs and strong adaptability.

[0113] Example 1: Pure electric passenger vehicle (basic version, without additional modules)

[0114] This embodiment targets ordinary pure electric passenger vehicles, adopting a configuration of basic mandatory data collection and no additional modules. The cloud uses a lightweight version of public cloud to achieve full-dimensional fault detection of the vehicle's mechanical, electrical, and basic network systems. It does not require core data from car manufacturers, has low implementation costs, and is suitable for the daily operation and maintenance needs of private cars.

[0115] 1. The vehicle-mounted edge detection unit implements hybrid...

[0116] Multi-source data acquisition module: Acquires CAN bus data (Battery Management System (BMS), Motor Controller (MCU), Vehicle Controller (VCU) data), original vehicle sensor data (battery temperature, voltage, current, motor temperature, speed, braking pressure, etc.), and basic image data from the original vehicle camera via the original vehicle's OBD interface and gateway, without requiring additional hardware. The acquisition frequency is adaptively adjusted to 1Hz~50Hz. Adaptive spatiotemporal alignment and preprocessing module: Performs hardware coarse synchronization using the original vehicle's T-Box as the main clock node, and software fine alignment using linear interpolation + Kalman filtering, mapping the data to a 10ms time axis. Outliers are removed using the 3σ criterion, and sensor electromagnetic interference is filtered out using adaptive wavelet threshold denoising, completing bus data packet loss. Lightweight local feature extraction module: Extracts frequency and time domain features from battery and motor time-series data using traditional signal processing; extracts basic image data from the original vehicle camera using a lightweight version of MobileNetV3. Spatial features are extracted to generate low-dimensional feature vectors; Adaptive edge inference engine: Adopting a multi-task learning lightweight architecture, the model parameters are adaptively adjusted based on vehicle mileage (30,000 km) and driving style (smooth) to infer battery data and determine "slight aging of battery cells (85% probability of failure, 10% uncertainty), remaining lifespan of approximately 2,500 km". An interpretable result is generated through an attention mechanism: "excessive battery voltage difference (65% contribution) + increased internal resistance (30% contribution), determined to be battery cell aging"; Hierarchical execution control module: This fault is a level two fault (warning level), with a pop-up warning on the local vehicle system. When the vehicle is connected to WiFi, the fault feature vector is uploaded to the cloud, and maintenance suggestions are pushed to the owner's mobile device; System self-diagnosis module: Real-time monitoring of the working status of each acquisition module, detection of data acquisition delay from the original vehicle vibration sensor, generation of system fault code, display of "sensor communication delay" on the dashboard, reminding the owner to check.

[0117] 2. Implementation of Cloud-based Big Data Analysis Center

[0118] Lightweight data receiving and parsing module: Receives battery fault feature vectors uploaded from the edge side, decompresses and decrypts them, and stores them in a classified manner to the public cloud lightweight version; Deep diagnosis and root cause localization module: Combines historical fault data of this vehicle model to correct the diagnostic results from the edge side, confirming "slight aging of individual battery cells" and the root cause as "normal degradation due to excessive battery charge-discharge cycles"; Interpretable lightweight fault knowledge graph module: Calls a basic general knowledge graph, uses the lightweight version of GNN for inference, generates the inference path "excessive charge-discharge cycles → increased internal resistance of individual battery cells → excessive voltage difference → aging of individual battery cells", and labels the confidence level of each step; Lightweight The system includes: a digital twin simulation module (using a basic experience-based twin, injecting battery aging parameters, extrapolating a remaining lifespan of approximately 2400 km ± 300 km, and cross-validating with edge-side results); a lightweight federated learning model iteration module (adding the fault case to the lightweight training set, aggregating gradient data from other vehicles, generating a new model version, and distributing it to the vehicle's edge side via OTA); and a remote operation and maintenance service module (pushing in-depth diagnostic results, remaining lifespan predictions, and maintenance recommendations ("Recommended battery equalization maintenance within 2000 km") to the owner's mobile device, while also recommending nearby new energy vehicle repair shops and supporting online appointments for owners).

[0119] 3. Implementation of lightweight communication transmission module

[0120] The edge device uploads fault feature vectors to the cloud via home WiFi, using lightweight SM4 encryption, with a transmission rate of 1Mbps and no data loss. The car owner's mobile device connects to the vehicle via Bluetooth to view fault information in real time and access maintenance suggestions and service station information pushed from the cloud via 5G.

[0121] Example 2: L2+ level autonomous driving pure electric minibus (advanced version, with added lightweight module)

[0122] This embodiment targets L2+ level autonomous pure electric minibuses, employing a configuration of basic mandatory data collection plus lightweight add-on modules. The cloud uses a lightweight expansion version of a private cloud to achieve full-stack fault detection of vehicle mechanics, electrical systems, software, connectivity, and network security, adapting to the high-precision and high-security operation and maintenance needs of fleets.

[0123] 1. Implementation of vehicle-mounted edge detection unit

[0124] Hybrid Multi-Source Data Acquisition Module: In addition to acquiring original vehicle data, it acquires high-frequency vibration sensor data (motor bearings, chassis suspension), micro gas sensor data (CO / H2 concentration in the battery pack), and network security detection sensor data (CAN bus message anomalies) via a plug-and-play lightweight add-on module, with the acquisition frequency adjusted to 100Hz~500Hz; Adaptive Spatiotemporal Alignment and Preprocessing Module: Based on hardware coarse synchronization, a simple GPS module is added to realize spatial location tagging, and software fine alignment is performed to form "spatiotemporal superframe" data; adaptive wavelet threshold denoising is used for high-frequency vibration data, and gas sensor data is smoothed; Lightweight Local Feature Extraction Module: Mini-LSTM is used to extract time-series features from the high-frequency vibration spectrum, and rate of change features are extracted from the gas sensor data, which are integrated with the original vehicle data features into a low-dimensional feature vector; Self-Adaptive Edge Inference Engine: The model parameters are adaptively adjusted according to the operating conditions of the autonomous minibus (urban roads, high-frequency start-stop). The system analyzes motor data and determines the problem as "early pitting corrosion of the motor bearing (90% probability of failure, 5% uncertainty), remaining lifespan approximately 1500 km). This can be interpreted as "abnormal high-frequency vibration characteristics of the motor bearing (80% contribution) + slight increase in motor temperature (15% contribution), leading to the determination of early pitting corrosion of the motor bearing." Simultaneously, abnormal messages are detected on the CAN bus, classified as "low-risk network security (intrusion of external messages, no impact)." The hierarchical control module handles the following: motor bearing failure is classified as a Level 3 fault (limited control level), triggering local audible and visual alarms, limiting the vehicle's maximum speed to 40 km / h, and immediately uploading to the cloud via 5G; network security failure is classified as a Level 1 fault (notification level), displaying a notification on the local dashboard without cloud upload; for abnormal CAN bus messages, a lightweight network security protection mechanism is triggered locally to block the abnormal messages, achieving local self-healing; the system self-diagnosis module monitors the working status of the added modules in real time, ensuring all modules operate normally without system fault codes.

[0125] 2. Implementation of Cloud-based Big Data Analysis Center

[0126] Lightweight data receiving and parsing module: Receives motor fault data uploaded from the edge via 5G high priority, decompresses and stores it in a lightweight expanded version of the private cloud; Deep diagnosis and root cause localization module: Combines the operating data of the autonomous minibus to confirm "early pitting corrosion of the motor bearing," with the root cause being "bearing fatigue caused by high-frequency start-stop in urban roads"; Explainable lightweight fault knowledge graph module: Calls the vehicle-specific knowledge graph (exclusive to autonomous minibuses), and generates the inference path "high-frequency start-stop → motor bearing load fluctuation → early pitting corrosion of the bearing → abnormal high-frequency vibration characteristics" through a lightweight version of GNN inference; Lightweight digital twin simulation module: Employs a high-fidelity mechanism-based twin (based on data provided by the vehicle manufacturer). The system simplifies motor design parameter construction, injects bearing pitting parameters, and simulates the remaining service life of approximately 1400 km ± 200 km. It also reproduces fault characteristics with a 95% fit to real-vehicle data. A lightweight federated learning model iteration module adds the fault case (a long-tail fault specific to autonomous minibuses) to the training set, fine-tunes the model using a few-shot learning algorithm, generates a sub-model specific to autonomous minibuses, and distributes it to all vehicles in the fleet via OTA. A remote maintenance and service module pushes fault results and speed limit suggestions to the fleet management platform and driver mobile devices, while simultaneously pushing fault information and maintenance suggestions ("Replace motor bearings immediately") to the fleet repair station, automatically generating maintenance work orders.

[0127] 3. Implementation of lightweight communication transmission module

[0128] The edge device transmits motor fault data via 5G with high priority, with a latency of ≤8ms. It uses the national cryptographic SM4 encryption + MQTT over TLS 1.3 lightweight protocol to ensure data security. The fleet management platform connects to the cloud via a private cloud to view the fault status of all vehicles in the fleet in real time and achieve unified management.

[0129] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A comprehensive intelligent fault detection system for automobiles, characterized in that: It is used to realize full-stack fault detection, root cause localization, prediction and early warning, remote operation and maintenance, and self-adaptive optimization for vehicle mechanical systems, electrical systems, software systems, connected systems and network security; It adopts a lightweight cloud-edge-device three-layer collaborative architecture, which includes an in-vehicle edge detection unit, a cloud big data analysis center, and a lightweight communication transmission module. The three components achieve data interaction and command issuance through a two-way encrypted communication link. The system integrates five core technologies: adaptive multi-source data fusion, interpretable lightweight knowledge graph, lightweight digital twin, lightweight federated learning, and self-adaptive model iteration. The overall design is modular and decoupled, supporting plug-in upgrades and functional expansion. It adopts a hybrid data acquisition mode of "original vehicle data as the main source and optional sensor data as the auxiliary source", which is compatible with traditional fuel vehicles, hybrid vehicles, pure electric vehicles and L2+ level and above autonomous vehicles.

2. The intelligent vehicle fault detection system according to claim 1, characterized in that: The vehicle-mounted edge detection unit is the core of the system for data acquisition, real-time inference, and local execution. It adopts a lightweight design of "original vehicle hardware reuse + optional add-on modules" and includes a hybrid multi-source data acquisition module, an adaptive spatiotemporal alignment and preprocessing module, a lightweight local feature extraction module, an adaptive edge inference engine, a hierarchical execution control module, and a system self-diagnosis module. Each module works together to complete data processing, local inference, fault execution, and system fault detection.

3. The intelligent vehicle fault detection system according to claim 2, characterized in that: The data collected by the hybrid multi-source data acquisition module is divided into basic mandatory data and advanced optional data. The basic mandatory data is obtained by reusing the original vehicle hardware, while the advanced optional data is obtained by external plug-and-play lightweight automotive-grade aftermarket module. The module supports multi-protocol compatibility, can parse original vehicle DBC files and ARXML files, and has reserved standardized sensor interfaces to support the access of third-party lightweight sensors.

4. The intelligent vehicle fault detection system according to claim 2, characterized in that: The adaptive spatiotemporal alignment and preprocessing module adopts a lightweight spatiotemporal alignment strategy of "hardware coarse synchronization + software fine alignment". It uses the original vehicle T-Box as the main clock node and maps the data to the global time axis through linear interpolation and Kalman filtering. At the same time, it uses lightweight algorithms such as the 3σ criterion and adaptive wavelet threshold denoising method to complete data cleaning, denoising, and completion, and performs lightweight compression and coarse feature extraction on unstructured data.

5. The intelligent vehicle fault detection system according to claim 2, characterized in that: The adaptive edge inference engine adopts a lightweight multi-task learning architecture, sharing the underlying feature extraction layer, and outputting fault type identification and classification tasks and remaining life prediction and regression tasks in parallel at the upper layer. It can adaptively adjust model parameters according to vehicle usage time, mileage, driving style, and usage conditions, introduce a lightweight attention mechanism to achieve interpretable inference, and complete the inference uncertainty assessment through lightweight Monte Carlo Dropout technology.

6. The intelligent vehicle fault detection system according to claim 2, characterized in that: The hierarchical execution control module adopts a dual hierarchical strategy of "fault severity + reasoning uncertainty" to classify faults into four levels: prompt level, early warning level, limited control level, and emergency level, and formulate corresponding execution strategies. It supports lightweight OTA self-healing, and can directly issue control commands at the edge side to complete reset or calibration for soft faults such as software logic errors and abnormal communication parameters.

7. The intelligent vehicle fault detection system according to claim 2, characterized in that: The system self-diagnosis module can monitor the working status of each module of the vehicle edge detection unit in real time. When an abnormality occurs, it generates a system fault code and pushes a prompt. When the inference engine is abnormal, it automatically switches to a backup lightweight model. When the network is interrupted, it automatically caches fault data and supports breakpoint resume. It also supports local self-testing and cloud-based remote self-testing and can generate a system health report.

8. The intelligent vehicle fault detection system according to claim 1, characterized in that: The cloud-based big data analysis center is deployed on a lightweight cloud server cluster, supporting on-demand deployment in public / private clouds. It serves as the core for in-depth system diagnosis, model iteration, knowledge graph updates, digital twin simulation, and remote operation and maintenance. It includes a lightweight data receiving and parsing module, a deep diagnosis and root cause localization module, an interpretable lightweight fault knowledge graph module, a lightweight digital twin simulation module, a lightweight federated learning model iteration module, and a remote operation and maintenance and service module. The computing power requirements of each module are reduced by more than 50% compared to existing technologies.

9. The intelligent vehicle fault detection system according to claim 8, characterized in that: The lightweight federated learning model iteration module adopts the "cloud global model + vehicle local small batch training" mode, only uploading gradient update values ​​instead of the original data, and completing lightweight gradient aggregation through weighted averaging. It can fine-tune the model using a few-shot learning algorithm for long-tail faults, generate lightweight sub-models for different vehicle models and working conditions, and distribute them to the edge via OTA, while retaining the old version model as a rollback backup.

10. The intelligent vehicle fault detection system according to claim 1, characterized in that: The lightweight communication transmission module adopts a "multi-protocol compatibility + lightweight optimization" design, supports multiple communication methods such as 5G / 4G / WiFi / Bluetooth, and adopts a multi-protocol layered communication mode of "wired communication inside the vehicle + wireless layered communication between the vehicle and the cloud". It adopts a transmission strategy of "feature vector as the main component and raw data as the auxiliary component", and optimizes data transmission by combining data fragmentation transmission and breakpoint resume mechanism. It ensures communication security through the national cryptographic SM4 lightweight encryption algorithm and the MQTT over TLS1.3 lightweight protocol, and supports bidirectional data interaction with mobile terminals.

Images