A Deep Learning-Based Method and System for Vehicle Fault Prediction and Remote Diagnosis

By combining the "edge-cloud" architecture with the lightweight TCN-ATTN model, the problems of insufficient model performance and limited remote diagnostic functions in vehicle fault diagnosis are solved, achieving efficient and accurate fault prediction and location, and improving vehicle operation safety and maintenance efficiency.

CN123065766APending Publication Date: 2026-11-10CHEMAI YUNSHU (BEIJING) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610986183.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-03
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing vehicle fault diagnosis technologies suffer from problems such as insufficient model performance, lack of interpretability, limited remote diagnostic functions, and low data processing efficiency, resulting in low vehicle fault identification accuracy, low maintenance efficiency, and high network latency.

Method used

The system employs a deep learning-based vehicle fault prediction and remote diagnosis system. It adopts a three-level distributed architecture of "end-edge-cloud" and combines a lightweight TCN-ATTN model and an adaptive sliding window feature extraction algorithm to achieve local fault prediction and hierarchical interpretable fault location, and supports full-function remote diagnosis.

Benefits of technology

It improves the accuracy of fault prediction and location, reduces maintenance and troubleshooting time, realizes efficient remote diagnostic functions and rapid response, and enhances vehicle operation safety and user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123065766A_ABST
    Figure CN123065766A_ABST
Patent Text Reader

Abstract

This invention provides a vehicle fault prediction and remote diagnosis method and system based on deep learning. The vehicle fault prediction and remote diagnosis system based on the improved TCN-ATTN hybrid model realizes distributed computing through a three-level "end-edge-cloud" architecture. Combined with a hierarchical interpretable fault location mechanism and full-function remote diagnosis based on UDS overIP, it significantly improves the accuracy, lead time and maintenance efficiency of fault prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of vehicle fault diagnosis technology, and in particular to a deep learning-based method and system for vehicle fault prediction and remote diagnosis, electronic device, and computer-readable storage medium. Background Technology

[0002] With the increasing electrification and intelligence of automobiles, modern cars are equipped with more than 100 electronic control units (ECUs) and hundreds of sensors to monitor the operating status of core components such as the engine, transmission, and braking system in real time. Traditional vehicle fault diagnosis mainly relies on a passive repair method of reading fault codes (DTCs) after a fault occurs. This method has a significant lag and often leads to sudden breakdowns of the vehicle while driving, which not only affects the user experience but may also cause serious traffic safety accidents.

[0003] Currently, data-driven vehicle fault prediction technology has gradually become a research hotspot, but existing technologies still have the following key shortcomings:

[0004] Insufficient model performance: Most models use traditional machine learning algorithms (such as SVM, random forest) or simple recurrent neural networks (RNN / LSTM), which are difficult to effectively capture long-range dependencies and complex nonlinear features in vehicle sensor time series data. The accuracy of identifying early minor faults is less than 70%, and the average prediction lead time is less than 10 minutes.

[0005] Lack of interpretability: Fault prediction results are mostly "black box" outputs, making it impossible to accurately locate faulty components. This forces maintenance personnel to conduct a comprehensive inspection for 2-3 hours, significantly reducing maintenance efficiency.

[0006] Remote diagnostic functions are limited: most can only perform simple fault code reading and warning push, and cannot support complex diagnostic operations such as active testing, freeze frame acquisition, and ECU parameter adjustment, making it difficult to meet the needs of modern automotive remote diagnostics.

[0007] Low data processing efficiency: The fixed sliding window feature extraction method cannot adapt to the feature change patterns of different fault types, resulting in insufficient feature extraction. At the same time, all computing tasks are concentrated on the cloud platform, resulting in high network latency and slow system response.

[0008] Therefore, there is an urgent need in this field to break through existing technological bottlenecks and develop a new generation of vehicle fault prediction systems with high accuracy, strong interpretability, multi-functional remote diagnostics and efficient data processing capabilities, so as to achieve early and accurate warning, rapid location and efficient repair of vehicle faults, ensure driving safety and improve user experience. Summary of the Invention

[0009] To address the technical problems existing in the prior art, the present invention provides the following technical solution:

[0010] On the one hand, a deep learning-based vehicle fault prediction and remote diagnosis system is provided, comprising an in-vehicle terminal layer, an edge computing layer, and a cloud platform layer, wherein:

[0011] The vehicle-mounted terminal layer is used to collect vehicle sensor data, perform local data preprocessing and local fault prediction, and execute remote diagnostic commands.

[0012] The edge computing layer is used to perform more accurate fault prediction and fault location on the early warning data sent by the vehicle terminal.

[0013] The cloud platform layer is used to store historical vehicle data, train and update fault prediction models, generate diagnostic reports, and interact with users.

[0014] Preferably, the vehicle terminal layer includes a data acquisition module, a local preprocessing module, a local prediction module, a remote diagnostic execution module, and a communication module. The data acquisition module is used to acquire standard parameters output by the vehicle's OBD interface and sensor data from each ECU in real time. The local preprocessing module is used to clean, normalize, and extract features from the acquired raw data. The local prediction module is used to perform real-time fault prediction on the vehicle terminal. The remote diagnostic execution module is used to receive and execute remote diagnostic commands issued by the cloud platform. The communication module is used to realize data communication between the vehicle terminal, the edge computing layer, and the cloud platform.

[0015] Preferably, the local preprocessing module uses an adaptive sliding window slicing algorithm to slice the time series data. The formula for the adaptive sliding window slicing algorithm is as follows:

[0016] in, Let be the size of the sliding window at time t. Based on the size of the sliding window, To adjust the coefficient, Let t be the rate of change of the fault characteristics within the window preceding time t. This represents the maximum characteristic change rate for this fault type.

[0017] Preferably, the local prediction module adopts a lightweight TCN-ATTN model, which includes an input layer, three causal convolutional layers, one multi-head self-attention layer, and two fully connected layers. The input layer has an input dimension of (60, 128), representing 128 sensor parameters from the past 60 minutes. The causal convolutional layers have a kernel size of 3, with dilation rates of 1, 2, and 4, and 64 output channels per layer. The multi-head self-attention layers have 4 heads and an attention dimension of 16. The fully connected layers have 128 hidden neurons and 3 output neurons, corresponding to the fault probability, fault type, and remaining available time, respectively.

[0018] Preferably, the edge computing layer includes a data receiving module, an edge prediction module, a fault location module, and a data forwarding module; the fault location module adopts a three-layer location architecture of "attention weight - feature contribution - component mapping", and combines the attention weight inside the model and the external SHAP value analysis to achieve layered fault location;

[0019] The formula for calculating the SHAP value is as follows:

[0020]

[0021] in, Let be the SHAP value of the i-th feature. For the set of all features, For a feature subset that does not contain the i-th feature, This is the model's predicted output for the feature subset S.

[0022] Preferably, the cloud platform layer includes a data storage module, a model training module, a global prediction module, a diagnostic report generation module, a remote diagnostic control module, and a user interaction module. The model training module uses a multi-task loss function to train the fault prediction model, and the formula for the multi-task loss function is: in, For binary classification loss function, For multi-class loss function, For regression loss function, and These are the weighting coefficients.

[0023] Preferably, the remote diagnostic execution module and the remote diagnostic control module implement remote diagnostic functions based on the UDSoverIP protocol, supporting services such as reading fault codes, reading frozen frame data, reading data streams, active testing, clearing fault codes, and adjusting ECU parameters.

[0024] Preferably, the system supports three operating modes: normal operation mode, early warning mode, and emergency mode. In normal operation mode, the vehicle terminal performs a local prediction every 5 minutes. In early warning mode, the edge computing layer performs a prediction every 1 minute. In emergency mode, the cloud platform immediately sends emergency early warning information and automatically issues remote diagnostic commands.

[0025] Preferably, the data acquisition module collects 128 standard OBD parameters, including engine speed, coolant temperature, intake pressure, oxygen sensor voltage, fuel pressure, and vehicle speed, at a frequency of 1Hz.

[0026] On the other hand, a deep learning-based method for vehicle fault prediction and remote diagnosis is provided, implemented based on the aforementioned system, including the following steps:

[0027] The vehicle-mounted terminal collects vehicle sensor data in real time, performs data cleaning, outlier processing, and adaptive sliding window feature extraction to generate feature vectors.

[0028] The vehicle terminal uses a lightweight TCN-ATTN model for local fault prediction. If a high-probability fault is predicted, the prediction results and related data are sent to the edge computing layer.

[0029] The edge computing layer uses the complete TCN-ATTN model for more accurate fault prediction and combines attention weights and SHAP values ​​for hierarchical fault localization.

[0030] The edge computing layer sends the fault prediction results and fault location reports to the cloud platform, and the cloud platform generates a detailed diagnostic report.

[0031] The cloud platform pushes fault warning information and diagnostic reports to vehicle owners and repair shops;

[0032] Repair shops or vehicle owners send remote diagnostic commands to the vehicle terminal through the cloud platform. The vehicle terminal executes the commands and returns the diagnostic results.

[0033] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0034] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0035] The beneficial effects of the technical solution provided by this invention include at least the following:

[0036] 1. Higher fault prediction performance: The TCN-ATTN hybrid model proposed in this invention achieves an accuracy of over 95% in identifying early and minor faults, with an average prediction lead time of over 30 minutes, which is far higher than the 70% accuracy and 10-minute lead time of existing technologies.

[0037] 2. Enhanced fault location capability: The hierarchical interpretable fault location mechanism proposed in this invention can accurately locate faulty components with an accuracy rate of over 92%, significantly reducing maintenance and troubleshooting time.

[0038] 3. More comprehensive remote diagnostic functions: Based on the UDSoverIP protocol, this invention realizes full-function remote diagnostics, supports 12 standard diagnostic services, and enables repair personnel to understand the vehicle's fault status before the vehicle arrives at the repair station, and prepare repair tools and parts in advance.

[0039] 4. More efficient system architecture: This invention adopts a three-level distributed architecture of "end-edge-cloud", which reasonably distributes computing tasks across each level, reducing the system response time from 5 seconds to 0.5 seconds, while reducing the computing load on the cloud platform by 70%.

[0040] 5. Improved adaptability and scalability: The adaptive sliding window feature extraction algorithm of this invention can automatically adjust the sliding window size according to the feature change patterns of different fault types, improving the targeting and effectiveness of feature extraction. Simultaneously, the system adopts a modular design, allowing for easy expansion to new fault types and new vehicle models. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is the core technology logic diagram of the present invention;

[0043] Figure 2 This is a schematic diagram of a system architecture provided by an embodiment of the present invention;

[0044] Figure 3 This is a block diagram of an edge server functional architecture provided by an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of a fault location mechanism provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of a cloud platform architecture provided in an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram illustrating an embodiment of a method provided in this invention. Detailed Implementation

[0048] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0049] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0050] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0051] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0052] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0053] I. Technical Research and Development Objectives of this Application

[0054] To address the shortcomings of the existing technologies, this invention proposes a vehicle fault prediction and remote diagnosis system based on an improved TCN-ATTN hybrid model. It achieves distributed computing through a three-level "end-edge-cloud" architecture, and combines a hierarchical interpretable fault location mechanism with full-function remote diagnosis based on UDSoverIP, which significantly improves the accuracy, lead time, and maintenance efficiency of fault prediction.

[0055] II. Technological Research and Development Innovation Points

[0056] (I) Explanation of System Improvements

[0057] 1. Improved TCN-ATTN hybrid fault prediction model: It organically combines the local feature extraction capability of causal convolution with the long-range dependency capture capability of multi-head self-attention mechanism, which solves the problem that traditional models cannot simultaneously take into account local details and global trends. The accuracy of early weak fault identification reaches more than 95%, and the average prediction lead exceeds 30 minutes.

[0058] 2. Layered interpretable fault localization mechanism: For the first time, a three-layer localization architecture of "attention weight - feature contribution - component mapping" is proposed. By combining the internal attention weight of the model and the external SHAP value analysis, the accurate localization from fault type to specific component is achieved, with a localization accuracy of over 92%.

[0059] 3. Full-featured remote diagnostics based on UDSoverIP: It implements 12 standard diagnostic services, including fault code reading, freeze frame acquisition, active testing, and ECU parameter adjustment. It supports multiple communication protocols such as CAN, CANFD, and Ethernet, and is compatible with more than 90% of mainstream automotive brands.

[0060] 4. Adaptive sliding window feature extraction algorithm: The sliding window size and step size are dynamically adjusted according to the rate of change of fault features. Compared with the fixed sliding window method, the feature extraction efficiency is improved by 40% and the model training time is shortened by 30%.

[0061] 5. "End-Edge-Cloud" three-level distributed architecture: The computing tasks are reasonably distributed across the vehicle terminal, the edge computing layer and the cloud platform, reducing the system response time from 5 seconds to 0.5 seconds, while reducing the computing load on the cloud platform by 70%.

[0062] (II) Core Technical Logic of the System

[0063] like Figure 1 As shown, the data acquisition module of the vehicle terminal acquires vehicle sensor data in real time at a frequency of 1Hz and sends it to the local preprocessing module.

[0064] The local preprocessing module cleans and extracts features from the data, and sends the feature vectors to the local prediction module.

[0065] The local prediction module performs real-time fault prediction. If a high-probability fault is predicted (>0.7), the prediction results and related data are sent to the edge computing layer.

[0066] The data receiving module of the edge computing layer receives data and sends it to the edge prediction module for more accurate prediction;

[0067] The edge prediction module sends the prediction results to the fault location module for fault location.

[0068] The fault location module generates a fault location report, which is sent to the cloud platform along with the prediction results.

[0069] The cloud platform's data storage module stores all data, while the diagnostic report generation module generates detailed diagnostic reports.

[0070] The cloud platform's user interaction module pushes fault warning information and diagnostic reports to vehicle owners and repair shops;

[0071] Repair shops or vehicle owners can send remote diagnostic commands to the vehicle terminal through the remote diagnostic control module of the cloud platform;

[0072] The remote diagnostic execution module of the vehicle terminal executes instructions and returns the diagnostic results to the cloud platform.

[0073] The application architecture and principles of this system will be explained in detail below.

[0074] III. System Architecture Introduction

[0075] like Figure 2 As shown, this system adopts a three-tier distributed architecture of "device-edge-cloud", with clearly defined functions at each level, enabling collaborative operation.

[0076] Vehicle terminal layer: responsible for real-time data acquisition, local preprocessing and lightweight fault prediction, and execution of remote diagnostic commands;

[0077] Edge computing layer: Deployed in 4S stores, service stations and regional data centers, responsible for accurate fault prediction and fault location of vehicles within the region;

[0078] Cloud platform layer: As the core of the system, it is responsible for global data storage, model training and updates, diagnostic report generation, and user interaction.

[0079] After diagnosis, the cloud platform can interact with the user's terminal and push diagnostic results, etc.

[0080] (a) Vehicle terminal layer (the hardware and control principles of the vehicle terminal can refer to the installation and application methods of existing vehicle-mounted infotainment systems)

[0081] The vehicle terminal layer is installed on the vehicle's OBD interface, is plug-and-play, requires no modification to the vehicle's wiring, and consists of the following 5 modules:

[0082] 1. Data Acquisition Module

[0083] This module is primarily used for real-time acquisition of vehicle OBD standard parameters and sensor data from various ECUs. It connects to the vehicle's CAN bus via the OBD-II interface and supports communication protocols such as ISO15765-4, SAE J1939, and ISO14229. The acquired parameters include 128 standard OBD parameters such as engine speed, coolant temperature, intake air pressure, oxygen sensor voltage, fuel pressure, and vehicle speed, with a acquisition frequency of 1Hz.

[0084] The data acquisition module sends a data request to each ECU once per second, receives sensor data returned by the ECU, and stores it in a local cache with timestamps. The local cache has a capacity of 1GB and can store approximately 12 days of historical data.

[0085] 2. Local preprocessing module

[0086] This module is primarily used for cleaning, normalizing, and extracting features from raw data to provide high-quality input for the prediction model. The main methods include the following:

[0087] Missing value handling: Linear interpolation is used to handle data points with ≤3 consecutive missing values. The formula is as follows:

[0088] in, For missing values, and For consecutive valid values, This corresponds to a timestamp. Data segments with more than 3 consecutive missing values ​​are marked as invalid data and discarded.

[0089] Outlier detection: A modified 3σ principle is adopted, taking into account the trend of the time series. The formula is as follows:

[0090] in, and Let be the mean and standard deviation of the data within the window preceding time t. Outliers are replaced with the average of the two consecutive valid values.

[0091] Adaptive sliding window slicing: The sliding window size is dynamically adjusted based on the rate of change of fault characteristics, using the following formula: in:

[0092] The sliding window size at time t (in seconds);

[0093] Sets the base sliding window size; the default value is 300 seconds (5 minutes).

[0094] The adjustment coefficient has a range of [0,1] and a default value of 0.5.

[0095] Let be the rate of change of the fault characteristics within the window preceding time t;

[0096] This represents the maximum characteristic change rate for this fault type.

[0097] Statistical feature extraction: Extract 10 statistical features from the time series data within each sliding window: mean, variance, standard deviation, maximum value, minimum value, peak value, kurtosis, skewness, root mean square, and zero-crossing rate (routine statistical calculations are sufficient).

[0098] The local preprocessing module processes the collected raw data every 5 minutes. First, it cleans up missing and outlier values. Then, it automatically adjusts the sliding window size based on the current vehicle operating status and historical fault records. Finally, it extracts statistical features and generates a feature vector with dimensions (128, 10).

[0099] Traditional fixed sliding window mechanisms (referring to the use of a fixed-size time window for data slicing and feature extraction in time series data processing) typically consist of four parts: fixed window size setting, data slicing, feature extraction, and result output. The limitation of fixed sliding windows is their inability to adapt to dynamic changes in data features. When data features change rapidly, fixed sliding windows may fail to capture key information in a timely manner; conversely, when data features change slowly, fixed sliding windows may introduce excessive redundant information, reducing the model's prediction accuracy and efficiency.

[0100] Compared to the adaptive sliding window mechanism described here:

[0101] Comparison indicators Fixed sliding window Adaptive sliding window Window size Fixed value (e.g., 300 seconds) Dynamic adjustment (100-600 seconds) Feature variation adaptation Weak, unable to adapt to rapid changes Strong, can be adjusted according to the characteristic rate of change Prediction accuracy Low accuracy, prone to losing key information High efficiency, enabling timely capture of key information. computational efficiency Stable, but may contain redundant calculations. Dynamic adjustment can reduce redundant calculations. Resource consumption Fixed, which may waste resources Dynamic optimization can save resources.

[0102] Therefore, the adaptive sliding window approach used here has the following technical advantages:

[0103] 1) Dynamic adaptability: The adaptive sliding window can dynamically adjust the window size according to the rate of change of data features, which can better adapt to the needs of different vehicle operating conditions and fault types.

[0104] 2) Higher prediction accuracy: By adjusting the window size in real time, the adaptive sliding window can capture changes in fault characteristics more promptly, improving the accuracy and lead time of fault prediction.

[0105] 3) Optimized computational efficiency: Adaptive sliding windows can reduce unnecessary computation and data processing while ensuring prediction accuracy, thereby improving the system's operating efficiency and response speed.

[0106] 4) Resource consumption optimization: By dynamically adjusting the window size, the adaptive sliding window can reasonably allocate computing resources according to actual needs, reduce unnecessary resource waste, and improve the overall performance of the system.

[0107] 5) Better robustness: Adaptive sliding windows can better handle noise and outliers in the data, improving the robustness and stability of the system.

[0108] 3. Local Prediction Module

[0109] Traditional fault prediction models typically employ rule-based expert systems or simple machine learning models, with the following application structure:

[0110] Input layer: Directly collects raw sensor data, which typically contains dozens to hundreds of parameters and has a high data dimensionality;

[0111] Feature extraction layer: Employs manually designed feature engineering methods, such as statistical features, time-domain features, and frequency-domain features, to extract key information from the data;

[0112] Model layer: Commonly used models include traditional machine learning models such as Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Tree (GBDT), or simple neural network models;

[0113] Output layer: Outputs the probability or type of failure, and usually uses a threshold judgment method for failure early warning;

[0114] In vehicle fault prediction, traditional models typically require transmitting a large amount of raw data to the cloud for processing, which presents problems such as large data transmission volume, high latency, and privacy risks. In addition, the models are highly complex and difficult to run in real time on vehicle terminals.

[0115] This module is primarily used for real-time fault prediction on the in-vehicle terminal, enabling rapid detection of potential faults. It employs an improved lightweight TCN-ATTN model, the structure of which and its application principles are as follows:

[0116] 1). Model Structure

[0117] Input layer: dimension (60, 128), which is 128 preprocessed sensor parameters from the past 60 minutes;

[0118] Causal convolutional layers: 3 causal convolutional layers with a kernel size of 3 and dilation rates of 1, 2, and 4 respectively. Each layer has 64 output channels, which can effectively capture long-term dependencies in time series data.

[0119] Multi-head self-attention layer: 1 layer of multi-head self-attention with 4 heads and 16 attention dimensions, which can adaptively focus on key information at different time steps;

[0120] Fully connected layer: 2 fully connected layers, with 128 neurons in the hidden layer and 3 neurons in the output layer, which output the fault probability, fault type and remaining available time, respectively.

[0121] Activation functions: ReLU (hidden layer), Sigmoid (failure probability), Softmax (failure type), and Linear (remaining available time).

[0122] 2). Model Optimization and Deployment

[0123] After the model is trained on the cloud platform, it is compressed using INT8 quantization technology. The compressed model size is about 2MB, which is only about 1 / 10 of the traditional model.

[0124] Meanwhile, through techniques such as channel pruning and knowledge distillation, the model structure is further optimized to reduce computational load and memory consumption;

[0125] The optimized model can run in real time on the ARM Cortex-A53 processor, with a single-frame inference time of less than 100ms, meeting the real-time requirements of the vehicle terminal.

[0126] The local prediction module receives the feature vector output by the local preprocessing module every 5 minutes and inputs it into the lightweight TCN-ATTN model for inference.

[0127] The model outputs three results: failure probability, failure type, and remaining available time. When the failure probability is greater than 0.7, a local warning is triggered.

[0128] The early warning information includes the fault type, fault probability, remaining available time, and relevant sensor data. At the same time, the prediction results and relevant data are sent to the edge computing layer for further analysis and processing.

[0129] Compared to traditional fault prediction models, the lightweight TCN-ATTN model used in this module has the following technical advantages:

[0130] 1) Lightweight design: The model size is only 2MB, which can run in real time on the vehicle terminal without relying on cloud computing, reducing data transmission volume and latency;

[0131] 2) High-precision prediction: By employing causal convolution and multi-head self-attention mechanisms, it can effectively capture long-term dependencies and key information in time series data, resulting in higher prediction accuracy;

[0132] 3) Multi-dimensional output: It not only outputs the failure probability, but also the failure type and remaining available time, providing richer information for failure early warning and maintenance decision-making;

[0133] 4) Strong real-time performance: The single-frame inference time is less than 100ms, which can process sensor data in real time and detect potential faults in a timely manner;

[0134] 5) Privacy protection: Data is processed locally on the terminal without being transmitted to the cloud, effectively protecting users' privacy data.

[0135] 4. Remote Diagnostic Execution Module

[0136] This module is primarily used to receive and execute remote diagnostic commands issued by the cloud platform and return the diagnostic results to the cloud platform. This module implements remote diagnostic functionality based on the UDS overIP protocol (ISO14229-5) and supports the following core services:

[0137] Read fault codes (DTCs);

[0138] Read the frozen frame data;

[0139] Read real-time data stream;

[0140] Active test execution;

[0141] Clear fault codes;

[0142] ECU parameter adjustment.

[0143] The remote diagnostic execution module establishes a TLS encrypted connection with the cloud platform via TCP / IP protocol, receives UDS diagnostic commands issued by the cloud platform, converts them into CAN bus messages and sends them to the corresponding ECU, and then converts the response messages returned by the ECU into IP data packets and sends them back to the cloud platform.

[0144] 5. Communication module

[0145] This module is primarily used to enable data communication between the vehicle-mounted terminal and the edge computing layer and cloud platform. It supports both 4G / 5G cellular networks and Wi-Fi, employs the MQTT protocol for data transmission, and has a QoS level of 1 to ensure reliable and low-latency data transmission.

[0146] (ii) Edge computing layer

[0147] The edge computing layer is deployed in 4S stores, service stations or regional data centers, and each node can cover more than 100,000 vehicles in the region.

[0148] like Figure 3 As shown, the edge server mainly consists of the following four modules:

[0149] 1. Data receiving module

[0150] This module is primarily used to receive fault warning information and related data sent by the vehicle-mounted terminal. In practical implementation, the EMQX MQTT Broker can be used, supporting simultaneous connections to more than 100,000 vehicle-mounted terminals with a message throughput of 100,000 messages per second.

[0151] 2. Edge prediction module

[0152] This module is primarily used for more accurate fault prediction of warning data sent by vehicle terminals. It adopts the same complete TCN-ATTN model as the cloud platform, but only uses vehicle data from the most recent 24 hours for inference. The model structure is consistent with the cloud platform model, but the number of parameters is reduced by 30%, and the single-frame inference time is less than 1 second.

[0153] 3. Fault location module

[0154] This module is primarily used for accurate localization of predicted faults and identification of faulty components. It employs a three-layer localization architecture: "attention weight - feature contribution - component mapping," as follows: Figure 4 As shown:

[0155] First layer: Attention weight analysis: Extracting the attention weight matrix of the multi-head self-attention layer in the TCN-ATTN model. ,in Let the time step be the time step size. Calculate each time step. (The average attention weights of the i-th layer at time t) :

[0156] Identify the top 10 time steps that contribute the most to the prediction results.

[0157] Second layer: SHAP value calculation: The marginal contribution of each sensor feature to the prediction result is calculated using the SHAP value. The formula is as follows: in:

[0158] Let be the SHAP value of the i-th feature;

[0159] The set of all features;

[0160] A feature subset that does not contain the i-th feature (represented as: );

[0161] This is the model's predicted output for the feature subset S.

[0162] The third layer: faulty component mapping: establish a mapping table between sensor features and faulty components, and infer the most likely faulty component based on the top 5 sensor features with the greatest contribution through a weighted voting method.

[0163] The fault location module receives the fault prediction results output by the edge prediction module. First, it calculates the attention weight to determine the key time steps. Then, it calculates the SHAP value of each sensor feature. Finally, it determines the faulty component based on the mapping table and generates a fault location report.

[0164] 4. Data forwarding module

[0165] This module uses the HTTP / 2 protocol to transmit data with the cloud platform, supporting data compression and resume functionality to ensure efficient and reliable data transmission. The module is primarily used to forward the processing results from the edge computing layer to the cloud platform for storage and further processing.

[0166] (III) Cloud Platform Layer

[0167] like Figure 5 As shown, the cloud platform is deployed on a mainstream public cloud platform, adopts a distributed architecture, and can elastically expand computing and storage resources. It is divided into an infrastructure layer, an interface layer, and a core service layer, and mainly consists of the following core service modules:

[0168] 1. Data storage module

[0169] This module is primarily used to store historical data, model parameters, diagnostic reports, and user information for all vehicles. It employs a hybrid storage architecture.

[0170] HDFS: Stores large-scale historical sensor data with a storage capacity of 1PB;

[0171] MySQL: Stores structured data (vehicle information, user information, diagnostic reports, etc.);

[0172] Redis: Stores cached data (real-time vehicle status, model parameters, etc.).

[0173] Data storage has a 5-year lifespan and supports data backup and recovery.

[0174] 2. Model Training Module

[0175] This module is mainly used to train and update fault prediction models to continuously improve model performance.

[0176] The model building process is as follows:

[0177] Dataset Construction: Historical fault data and normal operation data were extracted from the data storage module to construct the training dataset. The dataset contains operational data from 1 million vehicles, totaling 1 billion records, covering more than 500 common fault types.

[0178] Model training: The Adam optimizer was used with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The loss function used was a multi-task loss function, such as the following expression: in:

[0179] The binary classification loss function (whether a fault has occurred) is adopted, using binary cross-entropy loss;

[0180] For multi-class classification loss function (fault type), cross-entropy loss is adopted;

[0181] The regression loss function (remaining available time) is used, employing mean squared error loss.

[0182] and These are weighting coefficients, with values ​​of 1.0 and 0.5 respectively (based on calibration experiments or empirical values).

[0183] Model updates: Incremental learning is used to update the model monthly with newly collected data to ensure that the model can adapt to new vehicle models and fault types.

[0184] 3. Global Prediction Module

[0185] This module is primarily used for global fault prediction across all vehicles, identifying potential systemic faults. It employs a complete TCN-ATTN model to batch process historical data from all vehicles, predicting faults that may occur within the next 30 days. For predicted high-probability faults, an automatic warning process is triggered.

[0186] 4. Diagnostic Report Generation Module

[0187] This module is primarily used to generate detailed diagnostic reports based on fault prediction and fault location results. In this solution, a template-based approach is used to generate diagnostic reports quickly. The report content includes:

[0188] Basic vehicle information (license plate number, vehicle model, mileage, etc.);

[0189] Fault prediction results (fault probability, fault type, remaining available time);

[0190] Fault location results (faulty component, feature contribution);

[0191] Fault cause analysis;

[0192] Repair recommendations;

[0193] Estimated repair costs.

[0194] 5. Remote diagnostic control module

[0195] This module primarily utilizes the UDS over IP protocol to implement remote diagnostic control, supporting batch issuance of diagnostic commands and simultaneous remote diagnostics of multiple vehicles. It provides a RESTful API interface for easy integration with third-party systems. It is used to send remote diagnostic commands to the in-vehicle terminal and receive and process diagnostic results.

[0196] 6. User Interaction Module

[0197] This module is primarily used to provide user interfaces for car owners, repair shops, and car manufacturers. For example, it can provide user interfaces through web front-ends and mobile apps, and supports real-time message push, data visualization, and remote diagnostic control.

[0198] This system supports three operating modes, which automatically switch based on the failure probability (a threshold can be set by the user; an example is shown below):

[0199] Normal operating mode: When the failure probability output by the local prediction module is less than 0.3, the system is in normal operating mode. The vehicle terminal performs local prediction every 5 minutes, without sending data to the edge computing layer, and only stores the data in the local cache.

[0200] Warning Mode: When the fault probability output by the local prediction module is between 0.3 and 0.7, the system enters warning mode. The vehicle terminal sends data to the edge computing layer, which performs a prediction every minute, closely monitoring the fault development trend.

[0201] Emergency Mode: When the failure probability output by the local prediction module or the edge prediction module is greater than 0.7, the system enters emergency mode. The cloud platform immediately sends emergency warning information to the vehicle owner and nearby repair shops, and automatically issues diagnostic instructions to read frozen frames and critical data streams.

[0202] IV. Deep Learning-Based Vehicle Fault Prediction and Remote Diagnosis Methods

[0203] This method is implemented based on the aforementioned system. For example... Figure 6 As shown, the method includes the following steps:

[0204] Step 1: Data Acquisition and Preprocessing

[0205] This step, through data preprocessing, effectively removes noise and outliers, thereby improving data quality.

[0206] Specifically, the vehicle terminal's data acquisition module collects 128 standard OBD parameters of the vehicle in real time at a frequency of 1Hz. The collected raw data is first processed for missing values: linear interpolation is used to fill data points with ≤3 consecutive missing values, while data segments with >3 consecutive missing values ​​are discarded. Then, an improved 3σ principle is used to detect outliers, replacing them with the average of two consecutive valid values. Finally, an adaptive sliding window slicing algorithm is used to slice the time series data, extracting 10 statistical features within each window to generate a feature vector of dimension (128, 10). Therefore, the adaptive sliding window slicing algorithm can automatically adjust the sliding window size according to the changing patterns of fault characteristics, improving the targeting and effectiveness of feature extraction.

[0207] Step 2: Local Fault Prediction

[0208] The local preprocessing module sends the generated feature vectors to the local prediction module, which uses a lightweight TCN-ATTN model for inference. The model input is the feature vectors from the past 60 minutes, and the output is the probability of a fault occurring within the next 30 minutes, the fault type, and the remaining available time. When the fault probability > 0.7, a local warning is triggered, and the prediction results and related data are sent to the edge computing layer. Local fault prediction can be performed locally and in real-time on the vehicle terminal, without relying on the network, offering fast response and timely detection of potential faults.

[0209] Step 3: Precise Edge Prediction

[0210] The data receiving module of the edge computing layer receives the warning data sent by the vehicle terminal and sends it to the edge prediction module. The edge prediction module adopts the complete TCN-ATTN model and uses vehicle data from the most recent 24 hours for inference to obtain more accurate fault prediction results.

[0211] Edge computing layers leverage more powerful computing capabilities and more historical data to improve the accuracy of fault prediction while reducing the computing burden on cloud platforms.

[0212] Step 4: Layered Fault Location

[0213] The fault location module receives the fault prediction results output by the edge prediction module. First, it extracts the attention weights of the multi-head self-attention layer in the model, determining the top 10 time steps that contribute most to the prediction results. Then, it calculates the SHAP value for each sensor feature, quantifying the marginal contribution of each feature to the prediction result. Finally, based on the mapping table between sensor features and faulty components, it infers the most likely faulty component using a weighted voting method. This hierarchical fault location mechanism, combining attention weights and SHAP values, can accurately locate faulty components with an accuracy rate exceeding 92%, significantly reducing maintenance and troubleshooting time.

[0214] Step 5: Cloud platform processing and diagnostic report generation

[0215] The cloud platform can centrally manage all vehicle data, generate standardized diagnostic reports, and provide comprehensive fault information for repair shops and vehicle owners.

[0216] The edge computing layer sends the fault prediction results and fault location reports to the cloud platform, where the data storage module stores all the data. The diagnostic report generation module generates a detailed diagnostic report based on the fault prediction and fault location results, including basic vehicle information, fault prediction results, fault location results, fault cause analysis, repair recommendations, and estimated repair costs.

[0217] Step 6: Fault Alert Push Notification

[0218] The cloud platform's user interaction module pushes fault warnings and diagnostic reports to vehicle owners and nearby repair shops via SMS and app notifications. For emergency faults, the system automatically calls the vehicle owner to provide a voice alert. Timely fault warnings allow vehicle owners to take preventative measures, avoiding sudden breakdowns while driving and improving vehicle safety.

[0219] Step 7: Remote Diagnostic Execution

[0220] Repair shops or vehicle owners can send remote diagnostic commands to the vehicle terminal via the cloud platform's remote diagnostic control module, such as reading fault codes, reading freeze frames, and performing active tests. The vehicle terminal's remote diagnostic execution module receives and executes the commands, returning the diagnostic results to the cloud platform. The cloud platform then displays the diagnostic results on the user interface for repair personnel to reference. This remote diagnostic function allows repair personnel to understand the vehicle's fault condition before it arrives at the repair shop, enabling them to prepare repair tools and parts in advance and improving repair efficiency.

[0221] Example 1

[0222] This embodiment uses 10,000 SUV models of a certain car brand as the application object to deploy the vehicle fault prediction and remote diagnosis system of the present invention.

[0223] Vehicle-mounted terminals: 10,000 VT-200 OBD vehicle-mounted terminals are deployed, with one installed in each vehicle.

[0224] Edge computing layer: One edge computing layer is deployed in each of the three main 4S stores in the city. The server is configured with an Intel Xeon E5-2680v4 processor, 32GB of memory, and a 1TB SSD hard drive.

[0225] Cloud platform: Deployed on Alibaba Cloud ECS, configured with 2 core servers, 2 NVIDIA Tesla T4 GPUs, 10 application servers, and 5 database servers.

[0226] Taking an SUV with the license plate number Zhejiang A・xxxxx as an example, the specific implementation process of this system is described in detail

[0227] 1. Data collection and preprocessing: At 9:00 a.m. on October 1, 2025, the vehicle owner starts the vehicle, and the on-board terminal begins to collect 128 OBD standard parameters of the vehicle at a frequency of 1Hz. By 10:00 a.m., the on-board terminal has collected 3600 pieces of data. The local preprocessing module cleans these data, finds that there are 2 consecutive missing values in the voltage data of oxygen sensor 1, and fills them using the linear interpolation method; it finds that there is 1 abnormal value (12000rpm) in the engine speed data, which is replaced by the average value (2500rpm) of the two valid values before and after. Since the current vehicle is in normal operation, the change rate of fault features is small, and the sliding window size remains 300 seconds. 10 statistical features in each window are extracted to generate 12 feature vectors with dimensions of (128,10).

[0228] 2. Local fault prediction: At 10:00 a.m., the local prediction module receives the 12 feature vectors of the past 60 minutes and inputs them into the lightweight TCN-ATTN model for inference. The model output results are: fault probability 0.82, fault type P0300 (random misfire), and remaining available time 25 minutes. Since the fault probability is greater than 0.7, the on-board terminal triggers a local early warning, and sends the prediction result and related data to the nearest edge computing layer.

[0229] 3. Edge accurate prediction: The data receiving module of the edge computing layer receives the data at 10:00:02, and sends it to the edge prediction module. The edge prediction module uses the operation data of the vehicle in the last 24 hours for inference, and the output results are: fault probability 0.88, fault type P0300 (random misfire), and remaining available time 22 minutes.

[0230] 4. Hierarchical fault location: The fault location module receives the output result of the edge prediction module, first extracts the attention weight of the multi-head self-attention layer, and finds that the attention weights of the features of oxygen sensor 1 voltage, engine speed and intake pressure in the last 10 minutes are the highest. Then the SHAP value of each sensor feature is calculated, and the results show that the SHAP value of oxygen sensor 1 voltage is 0.35, the SHAP value of engine speed is 0.22, and the SHAP value of intake pressure is 0.18, and the total contribution of these three features reaches 75%. Finally, according to the mapping relation table, it is inferred that the most likely faulty components are the spark plug and fuel injector of the third cylinder.

[0231] 5. Cloud Platform Processing and Diagnostic Report Generation: At 10:00:05, the edge computing layer sends the fault prediction results and fault location report to the cloud platform. The cloud platform's data storage module stores all data, and the diagnostic report generation module generates a detailed diagnostic report, including:

[0232] Basic vehicle information: License plate number Zhejiang A·12345, vehicle type SUV, mileage 52,000 kilometers;

[0233] Fault prediction result: There is an 88% probability of a P0300 random misfire fault occurring within the next 22 minutes;

[0234] Fault location results: The most likely faulty components are the spark plug and fuel injector in cylinder 3;

[0235] Fault cause analysis: Incomplete fuel combustion caused by aging spark plugs or clogged fuel injectors;

[0236] Repair suggestion: Stop the vehicle immediately and contact the nearest repair shop to replace the spark plugs and clean the fuel injectors;

[0237] Estimated repair cost: approximately 800 yuan.

[0238] 6. Fault Warning Push: At 10:00:07, the cloud platform's user interaction module pushes fault warning information and diagnostic reports to the vehicle owner via the App and sends an SMS reminder. Simultaneously, the warning information is pushed to the nearest 4S store (within 2 kilometers). The 4S store's repair personnel receive the warning information at 10:00:10 and immediately prepare repair tools and parts.

[0239] 7. Remote Diagnostic Execution: The maintenance personnel send a command to the vehicle terminal via the remote diagnostic control module on the cloud platform to read fault codes and freeze frames. The remote diagnostic execution module on the vehicle terminal receives the command at 10:00:12, executes it, and returns the results to the cloud platform. The diagnostic results show: no fault codes are currently present, but the freeze frame data shows an engine speed of 2500 rpm, a coolant temperature of 90℃, and an oxygen sensor 1 voltage of 0.1V, consistent with the fault prediction results.

[0240] 8. Repair Process: After receiving the warning message, the car owner immediately drove the vehicle to the nearest 4S shop. Based on the diagnostic report and remote diagnostic results, the repair personnel directly inspected the spark plug and fuel injector of cylinder 3, finding that the spark plug was severely worn and the fuel injector was slightly clogged. The repair personnel replaced the spark plug and cleaned the fuel injector, resolving the problem. The entire repair process took only 30 minutes.

[0241] In this embodiment, the system successfully predicted an impending random fire fault and accurately located the faulty component. Compared with traditional passive maintenance methods, this system has the following significant advantages:

[0242] Predictive lead time: Faults are predicted 22 minutes in advance, preventing vehicles from suddenly breaking down while driving;

[0243] Location accuracy: The fault location result is completely consistent with the actual fault, and maintenance personnel do not need to perform any troubleshooting work;

[0244] Repair efficiency: Repair time was reduced from 2.5 hours to 0.5 hours, an improvement of 80%;

[0245] User experience: Timely warnings and accurate repair suggestions have significantly improved user satisfaction.

[0246] During the six-month trial operation, the system predicted 128 potential faults with a prediction accuracy of 95%, a fault location accuracy of 92%, an average repair time of about 0.5 hours, and user satisfaction increased from 82% to 96%.

[0247] Comparative Example 1

[0248] The comparison uses traditional vehicle fault diagnosis methods. The data comes from the maintenance records of a certain car brand's 4S store from January to June 2026 (for data testing only and no commercial use; both parties have signed a technology upgrade agreement), involving maintenance cases of 10,000 SUV models.

[0249] Traditional fault diagnosis process:

[0250] 1. Fault Reporting: After discovering a malfunction in their vehicle, car owners can call the 4S dealership's repair hotline or go directly to the dealership for repairs. For example, an SUV with license plate number Zhejiang A・xxxxx might experience engine vibration and insufficient power while driving, prompting the owner to visit the 4S dealership for repairs.

[0251] 2. Manual Diagnosis: The repair personnel first inquire about the vehicle owner's symptoms, and then make a preliminary judgment on the scope of the fault through visual inspection, auscultation, and other methods. If the repair personnel find that the engine idles unstablely and occasionally emits black smoke from the exhaust pipe, the preliminary judgment is that it may be a problem with the ignition system or fuel system.

[0252] 3. Fault Code Reading: Connect a car diagnostic tool to the vehicle's OBD interface and read the fault codes. The diagnostic tool displays fault code P0300 (random misfire), indicating that there is a misfire in the engine, but the specific cylinder and component with the fault cannot be determined.

[0253] 4. Instrument Testing: The technicians used a multimeter to test the ignition coil resistance and spark plug gap, a cylinder pressure gauge to measure the pressure in each cylinder, and a fuel pressure gauge to test the fuel system pressure. The test results showed that the pressure in cylinder 3 was 7.5 bar (normal range is 8-12 bar), which was below normal; the pressures in the other cylinders were normal; the ignition coil resistance and spark plug gap were within the normal range; and the fuel system pressure was normal.

[0254] 5. Fault Location: Based on the test results, the technicians suspected a clogged fuel injector or a faulty spark plug in cylinder 3. To pinpoint the specific faulty component, they used a substitution method, swapping the fuel injector and spark plug of cylinder 3 with those of cylinder 1, and then started the engine to observe if the fault shifted. After the swap, the fault persisted. Further inspection revealed that the cylinder pressure in cylinder 3 was still low, leading them to suspect a cylinder sealing problem.

[0255] 6. Disassembly and inspection: The repairman disassembled the engine cylinder head and found that the valve of cylinder 3 had serious carbon buildup, which caused the valve to not close properly and the cylinder pressure to be insufficient.

[0256] 7. Repair: The technicians cleaned the carbon deposits on the valves, reinstalled the cylinder head, and then started the engine for testing. The engine vibration disappeared, and the power returned to normal. The entire repair process took 2.5 hours.

[0257] During the 6-month statistical period, traditional fault diagnosis methods handled 120 engine misfires, with an average repair time of 2.5 hours, a fault location accuracy rate of 75%, and a user satisfaction rate of 82%.

[0258] In comparison:

[0259] Comparison indicators This embodiment Comparative Example Fault prediction lead time Prediction 22 minutes in advance No prior warning; repairs are only reported after the fault occurs. Fault location accuracy 92% 75% Mean Time Between Repairs 0.5 hours 2.5 hours User satisfaction 96% 82% Fault prediction accuracy 95% - Improved maintenance efficiency Increase by 80% -

[0260] Therefore, it can be seen that the vehicle fault prediction and remote diagnostic system of this embodiment has the following significant advantages over traditional fault diagnosis methods:

[0261] 1. Proactive prediction and prevention: This system can predict potential faults up to 22 minutes before they occur, preventing sudden breakdowns during driving and improving driving safety. Traditional methods, on the other hand, can only diagnose and repair faults after they occur, and cannot prevent them in advance.

[0262] 2. Precise Location, Efficient Repair: This system, through a lightweight TCN-ATTN model and SHAP value analysis, can accurately locate faulty components, achieving a fault location accuracy rate of 92%. Repair personnel can directly perform repairs based on the diagnostic results, eliminating the need for tedious troubleshooting. The average repair time is reduced from 2.5 hours to 0.5 hours, improving repair efficiency by 80%.

[0263] 3. Layered Diagnosis and Collaborative Processing: This system adopts a layered architecture consisting of an in-vehicle terminal, an edge computing layer, and a cloud platform, enabling collaborative processing of rapid local prediction, precise edge diagnosis, and comprehensive analysis by the cloud platform. This architecture ensures both real-time fault diagnosis and leverages the powerful computing capabilities of the cloud platform for in-depth analysis.

[0264] 4. Remote Service and Intelligent Push Notifications: This system can push fault warnings and diagnostic reports to vehicle owners via App and SMS, and also send warnings to the nearest 4S dealership. Repair personnel can prepare repair tools and parts in advance, improving repair response speed.

[0265] 5. Data-driven, continuous optimization: This system can collect and store vehicle operating data and maintenance records, and continuously optimize the fault prediction model through big data analysis to improve the accuracy and reliability of fault diagnosis.

[0266] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0267] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0268] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0269] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0270] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0271] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0272] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units 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 device, 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 devices or units may be electrical, mechanical, or other forms.

[0273] The units described 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.

[0274] In addition, the functional units in the various embodiments of the present invention 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.

[0275] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a 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 several 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 described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0276] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A vehicle fault prediction and remote diagnosis system based on deep learning, characterized in that, It includes the vehicle terminal layer, the edge computing layer, and the cloud platform layer, among which: The vehicle-mounted terminal layer is used to collect vehicle sensor data, perform local data preprocessing and local fault prediction, and execute remote diagnostic commands. The edge computing layer is used to perform more accurate fault prediction and fault location on the early warning data sent by the vehicle terminal. The cloud platform layer is used to store historical vehicle data, train and update fault prediction models, generate diagnostic reports, and interact with users.

2. The system according to claim 1, characterized in that, The vehicle terminal layer includes a data acquisition module, a local preprocessing module, a local prediction module, a remote diagnostic execution module, and a communication module; the data acquisition module is used to acquire standard parameters output by the vehicle's OBD interface and sensor data from each ECU in real time. The local preprocessing module is used to clean, normalize, and extract features from the collected raw data; the local prediction module is used to perform real-time fault prediction on the vehicle terminal. The remote diagnostic execution module is used to receive and execute remote diagnostic commands issued by the cloud platform; the communication module is used to realize data communication between the vehicle terminal and the edge computing layer and the cloud platform.

3. The system according to claim 2, characterized in that, The local preprocessing module uses an adaptive sliding window slicing algorithm to slice the time series data. The formula for the adaptive sliding window slicing algorithm is as follows: in, Let be the size of the sliding window at time t. Based on the size of the sliding window, To adjust the coefficient, Let t be the rate of change of the fault characteristics within the window preceding time t. This represents the maximum characteristic change rate for this fault type.

4. The system according to claim 2, characterized in that, The local prediction module employs a lightweight TCN-ATTN model, which includes: (1) Input layer, configured to receive multiple preprocessed sensor parameters from past time; (2) Causal convolutional layer, which contains multiple causal convolutions with an exponentially increasing kernel expansion rate, is used to capture long-term dependencies in time series data; (3) Multi-head self-attention layer, used to adaptively focus on key information at different time steps; (4) Fully connected layer, containing 3 output layer neurons, which output the fault probability, fault type and remaining available time respectively; The model is lightweighted using INT8 quantization, channel pruning, or knowledge distillation; Each time, the feature vector output by the local preprocessing module is received and input into the lightweight TCN-ATTN model for inference. When the fault probability is >0.7, a local early warning is triggered, and the prediction results and related data are sent to the edge computing layer for further analysis and processing.

5. The system according to claim 1, characterized in that, The edge computing layer includes a data receiving module, an edge prediction module, a fault location module, and a data forwarding module. The fault location module adopts a three-layer location architecture of "attention weight - feature contribution - component mapping", which combines the attention weight inside the model and the external SHAP value analysis to achieve layered fault location. The formula for calculating the SHAP value is as follows: in, Let be the SHAP value of the i-th feature. For the set of all features, For a feature subset that does not contain the i-th feature, This is the model's predicted output for the feature subset S.

6. The system according to claim 1, characterized in that, The cloud platform layer includes a data storage module, a model training module, a global prediction module, a diagnostic report generation module, a remote diagnostic control module, and a user interaction module. The model training module uses a multi-task loss function to train the fault prediction model. The formula for the multi-task loss function is: in, For binary classification loss function, For multi-class loss function, For regression loss function, and These are the weighting coefficients.

7. The system according to claim 6, characterized in that, The remote diagnostic execution module and remote diagnostic control module implement remote diagnostic functions based on the UDS overIP protocol, supporting functions such as reading fault codes, reading frozen frame data, reading data streams, active testing, clearing fault codes, and ECU parameter adjustment.

8. The system according to claim 1, characterized in that, The system supports three operating modes: normal operation mode, early warning mode, and emergency mode; in normal operation mode, the vehicle terminal performs a local prediction every 5 minutes. In early warning mode, the edge computing layer makes a prediction every minute; in emergency mode, the cloud platform immediately sends emergency warning information and automatically issues remote diagnostic instructions.

9. The system according to claim 1, characterized in that, The data acquisition module collects 128 standard OBD parameters, including engine speed, coolant temperature, intake pressure, oxygen sensor voltage, fuel pressure, and vehicle speed, at a frequency of 1Hz.

10. A method for vehicle fault prediction and remote diagnosis based on deep learning, implemented based on the system described in any one of claims 1-9, characterized in that, Includes the following steps: The vehicle-mounted terminal collects vehicle sensor data in real time, performs data cleaning, outlier processing, and adaptive sliding window feature extraction to generate feature vectors. The vehicle terminal uses a lightweight TCN-ATTN model for local fault prediction. If a high-probability fault is predicted, the prediction results and related data are sent to the edge computing layer. The edge computing layer uses the complete TCN-ATTN model for more accurate fault prediction and combines attention weights and SHAP values ​​for hierarchical fault localization. The edge computing layer sends the fault prediction results and fault location reports to the cloud platform, and the cloud platform generates a detailed diagnostic report. The cloud platform pushes fault warning information and diagnostic reports to vehicle owners and repair shops; Repair shops or vehicle owners send remote diagnostic commands to the vehicle terminal through the cloud platform. The vehicle terminal executes the commands and returns the diagnostic results.