New energy vehicle fault diagnosis method and system based on cloud edge collaboration

Abstract

The application discloses a new energy vehicle fault diagnosis method and system based on cloud edge cooperation, which comprises the following steps: collecting operation and environment data by a vehicle-mounted edge node, carrying out cleaning and double-scale feature extraction, and completing preliminary screening of faults through three-level threshold rules to generate candidate faults; combining candidate faults according to time windows to form fault events, calculating dynamic risk values by combining coupling correlation, deterioration trend, working condition, and four types of correction coefficients after static risk scoring, and dividing fatal, early warning, and healthy three-level fault grades. The system is hierarchically dispatched according to the grades: fatal grade faults are executed in a closed loop locally, early warning grade faults are deeply diagnosed by regional edge nodes, environment false alarm suppression is realized through multi-vehicle space-time aggregation and roadside data, and batch defect identification and whole life cycle health assessment are carried out on the healthy grade faults uploaded to the cloud. The application considers both safety response timeliness and diagnosis accuracy, can significantly reduce the false alarm rate, and improves the batch quality control ability.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and intelligent operation and maintenance technology for new energy vehicles, specifically a fault diagnosis method and system for new energy vehicles based on cloud-edge collaboration. Background Technology

[0002] With the rapid development of the new energy vehicle industry, vehicle power systems are becoming increasingly complex, encompassing multiple subsystems such as battery management system (BMS), motor control system, vehicle control unit (VCU), thermal management system, and high-voltage electrical system. These systems are highly coupled and have diverse fault modes, posing a severe challenge to traditional fault diagnosis technologies.

[0003] Existing fault diagnosis systems for new energy vehicles generally suffer from multiple technical problems, including insufficient classification accuracy, poor architecture adaptability, high false alarm rate, delayed batch risk detection, and imbalance in resource consumption. 1. Relying solely on a single-parameter static threshold for fault classification cannot take into account the effects of multi-parameter coupling, degradation evolution trends, real-time operating condition differences, and component batch quality attributes on the actual risk. This can easily lead to inaccurate classification, such as underestimating high-risk coupled faults and misjudging instantaneous parameter fluctuations, making it difficult to guarantee the accuracy of safety response. 2. The diagnostic architecture mostly adopts a centralized diagnostic mode of pure vehicle-side local diagnosis or full data upload to the cloud. The former is limited by the vehicle's computing power and is difficult to achieve in-depth diagnosis, while the latter is highly dependent on network transmission and has significant latency. It cannot simultaneously meet the real-time requirements of safety faults and the in-depth diagnostic needs of complex faults. 3. The lack of a regional-level multi-vehicle group perception and roadside multi-source data linkage mechanism makes it impossible to effectively distinguish between vehicle malfunctions and abnormal signals caused by external environmental interference such as road and weather conditions. The environmental false alarm rate remains high, resulting in high ineffective after-sales maintenance costs. 4. Batch-related component quality defects rely on passive backtracking of after-sales maintenance data, making it difficult to complete cluster identification and risk warning in the early stages of fault spread, resulting in lagging quality control; 5. The mode of repeatedly calculating feature data and uploading full data in real time will result in a large amount of redundant consumption of computing power and communication bandwidth, and it will also be difficult to support the implementation of full life cycle health assessment and predictive maintenance of core components. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a fault diagnosis method for new energy vehicles based on cloud-edge collaboration, comprising the following steps: Step s1: Collect the operating parameters and environmental data of new energy vehicles by deploying on-board edge nodes on the new energy vehicles, and perform data cleaning, feature extraction and preliminary fault judgment on the operating parameters; Step s2: When candidate faults are generated during the initial fault assessment process, all candidate faults are associated and merged to generate one or more fault events. Based on the preset static risk value scoring model, the static risk value of each fault event is output. Step s3: Obtain the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient. Perform a second correction on the static risk value based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient to obtain the dynamic risk value. Determine the fault level of the fault event based on the dynamic risk value. Step s4: Based on the regional scheduling strategy, critical fault events are assigned to the vehicle edge node for local execution, early warning fault events are assigned to the regional edge node for execution, and health-level fault events are uploaded to the cloud diagnostic center for execution. Step s5: Perform single-vehicle early warning level fault diagnosis on the complete data packets received by the regional edge nodes, generate fault prediction results, and perform environmental correlation discrimination and false alarm suppression operations based on the fault prediction results; Step s6: Perform batch defect clustering identification on the fault prediction results uploaded to the cloud diagnostic center from the regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the health-level fault events uploaded to the cloud diagnostic center to generate health-level fault diagnosis results.

[0005] Furthermore, the process of data cleaning and feature extraction includes: Time base alignment is performed on each type of parameter in the running parameters. After time base alignment, physical range and logical consistency verification, statistical outlier removal, multimodal noise filtering, and hierarchical missing data completion are performed on each type of parameter to obtain the data-cleaned parameters. A dual-scale sliding window is set up, and the scale sliding window corresponding to each type of parameter is divided according to the process characteristics of each type of parameter. The dual-scale sliding window includes a fast-changing window and a slow-changing window. Feature extraction is performed on each type of parameter that has completed data cleaning within different scale windows to obtain five types of features, including time domain features, frequency domain features, time-frequency domain features, degradation trend features, and system coupling correlation features.

[0006] Furthermore, the process of initial fault diagnosis includes: The system includes a pre-defined hard threshold rule base, a dynamic threshold rule base, and a health-level rule base. It compares the temporal statistical characteristics of each type of parameter with the hard threshold corresponding to the temporal statistical characteristics of each type of parameter in the hard threshold rule base. When the temporal statistical characteristics of a certain type of parameter reach the hard threshold within two consecutive corresponding scale sliding windows, the certain type of parameter is marked as a fatal candidate fault. The dynamic threshold library includes warning thresholds corresponding to the time-domain statistical characteristics of various types of parameters under different operating conditions. The operating conditions of new energy vehicles are obtained based on environmental data. The warning thresholds in the dynamic threshold library are called according to the current operating conditions. The time-domain statistical characteristics of each type of parameter are compared with the warning thresholds corresponding to the time-domain statistical characteristics of each type of parameter. When the time-domain statistical characteristics of a certain type of parameter reach the warning threshold within two consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the warning level. The degradation trend characteristics of each type of parameter are compared with the thresholds corresponding to the degradation trend characteristics of each type of parameter in the health-level rule base. When the degradation trend characteristics of a certain type of parameter reach the threshold within three consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the health level.

[0007] Furthermore, when candidate faults are generated during the initial fault assessment process, all candidate faults are correlated and merged to generate one or more fault events. The process of outputting the static risk value of each fault event based on a preset static risk value scoring model includes: When any candidate fault is generated, a merging time window is constructed with the generation timestamp of the candidate fault as the center. All generated candidate faults within the merging time window are merged into the merging candidate pool. Based on the preset vehicle fault domain classification system, the candidate faults in the merging candidate pool are clustered to generate one or more fault events. The primary anomaly and associated secondary anomalies in each fault event are determined based on the initial judgment level of each candidate fault in each fault event, and the scoring priority of each fault event is determined based on the initial judgment level of the primary anomaly in each fault event. A static risk value scoring model is preset. According to the scoring priority, the main abnormal event and the associated secondary abnormal data packets in each fault event are input into the static risk value scoring model. The static risk value of each fault event is output according to the static risk value scoring model. The data packet includes time-domain statistical characteristics, degradation trend characteristics, operating conditions, and preliminary assessment level.

[0008] Furthermore, the process of obtaining coupling correction coefficients, degradation trend correction coefficients, operating condition correction coefficients, and batch quality correction coefficients, and then performing a secondary correction on the static risk value based on these coefficients to obtain the dynamic risk value, and determining the fault level of the fault event based on the dynamic risk value includes: A fault parameter coupling association graph is preset. The type parameters corresponding to the main abnormal event and the associated secondary abnormal event in the fault event are used as nodes. The connection path between nodes is searched in the fault parameter coupling association graph. The coupling strength between nodes is determined according to the path length and edge attributes. The coupling correction coefficient is obtained according to the coupling strength between nodes. Extract the historical time-domain statistical features of the type parameters corresponding to the main abnormal events within the first k fast-changing windows of the merged time window, perform working condition normalization on the historical time-domain statistical features within each fast-changing window, obtain the relative degradation degree of each fast-changing window, perform linear fitting on the relative degradation degree of each fast-changing window, obtain the degradation rate of the main abnormal event, perform normalization on the degradation rate, and obtain the degradation trend correction coefficient based on the degradation rate after normalization. A preset working condition correction coefficient mapping table is provided, which includes working condition correction coefficients corresponding to different working conditions. The working condition correction coefficient matching the current working condition is obtained based on the working condition correction coefficient mapping table. Obtain the batch ID of the specific component corresponding to the main abnormal event in the fault event, obtain the batch quality level corresponding to the batch ID of the specific component from the cloud diagnostic center, and obtain the batch quality correction coefficient based on the batch quality level. The static risk value is corrected based on the coupling correction coefficient, the degradation trend correction coefficient, the operating condition correction coefficient, and the batch quality correction coefficient to obtain the dynamic risk value. Within the dynamic risk value range, threshold points are selected to divide the fault into sub-intervals of different fault levels. The fault level of the fault event is determined according to the sub-interval to which the dynamic risk value belongs. The fault levels include fatal, warning, and healthy levels.

[0009] Furthermore, based on the regional scheduling strategy, the process of allocating fatal fault events to onboard edge nodes for local execution, early warning fault events to regional edge nodes for execution, and health-level fault events to the cloud diagnostic center for execution includes: For critical fault events, the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the critical fault event are input into the local fault diagnosis model deployed on the vehicle edge node. The fault judgment result is output according to the local fault diagnosis model. The protection control command is generated according to the fault judgment result, and the fault event summary of the fault judgment result is extracted and sent to the regional edge node and cloud diagnosis center currently connected to the new energy vehicle. For warning-level fault events, the main abnormal event and the associated complete data packet of the warning-level fault event are uploaded to the edge node of the area currently connected to the new energy vehicle. The complete data packet includes five types of features, operating conditions, vehicle desensitization identifier and GPS location stamp. For health-level fault events, based on the operating conditions of new energy vehicles, the main abnormal events and associated five types of abnormal characteristics and operating conditions in the health-level fault events are uploaded to the cloud diagnostic center.

[0010] Furthermore, the process of performing single-vehicle early warning-level fault diagnosis on the complete data packets received by the regional edge nodes, generating fault prediction results, and performing environmental correlation discrimination and false alarm suppression operations based on the fault prediction results includes: When the edge node receives the complete data packet, it extracts the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the data packet, as well as the associated anomalies. It generates a multi-dimensional joint feature vector and inputs the multi-dimensional joint feature vector into the multi-scale attention diagnosis model deployed on the edge node. The multi-scale attention diagnosis model outputs the fault prediction result, which includes specific components and fault modes. The fault prediction result is then bound to the vehicle desensitization identifier, GPS location stamp and operating condition in the complete data packet. The fault prediction results corresponding to each vehicle's desensitized identifier are screened for fault modes. If the fault prediction result corresponding to the vehicle's desensitized identifier does not meet the fault mode requirements, a temporary handling suggestion is generated based on the fault prediction result corresponding to the vehicle's desensitized identifier and the operating conditions, and fed back to the corresponding vehicle edge node. If the fault prediction result meets the fault mode requirements, the following steps are executed: Step 1: Spatially divide the coverage area of ​​the regional edge node into several road segment units. Divide the complete data packets uploaded to the regional edge node into time-divided data packets in different time windows. Bind the fault prediction results corresponding to the complete data packets to the time window according to the time window to which the complete data packets belong. Step 2: Cluster the fault prediction results corresponding to the desensitized labels of each vehicle within the same time window and the same road segment unit, group the same fault prediction results into the same group, and obtain the number of fault prediction results in each group. If the number of fault prediction results in a certain group is greater than the preset warning threshold for the same type, then proceed to Step 3. If the number of fault prediction results in each group is less than or equal to the preset warning threshold for the same type, then proceed to Step 5. Step 3: Obtain roadside monitoring data for a specific road segment unit. Generate an environmental cause quantitative score based on the roadside monitoring data and fault prediction results. If the environmental cause quantitative score is greater than the preset environmental cause quantitative threshold, perform multi-vehicle coverage determination for the specific group. If the specific group is determined to have no batch clustering, proceed to Step 4. If the specific group is determined to have batch clustering, upload the fault prediction results of the specific group to the cloud diagnostic center to perform batch defect clustering identification. If the environmental cause quantitative score is less than or equal to the preset environmental cause quantitative threshold, proceed to Step 5. Step 4: Obtain the fault prediction results corresponding to the desensitized labels of each vehicle in a certain group; Step 5: For each vehicle desensitized label in each group, based on the fault prediction results and operating conditions corresponding to the vehicle desensitized label, generate temporary handling suggestions for the vehicle desensitized label and feed them back to the corresponding vehicle edge node.

[0011] Furthermore, the process of performing batch defect clustering identification on the fault prediction results uploaded by the regional edge nodes and generating batch quality levels corresponding to different batch IDs includes: Based on the fault prediction results corresponding to the vehicle desensitization labels in a certain group, the specific components and fault modes are determined. The batch IDs corresponding to the specific components are extracted from the vehicle desensitization labels. The specific components, the batch IDs corresponding to the specific components, and the fault modes are uploaded to the cloud diagnostic center. The cloud-based diagnostic center obtains the global baseline failure rate corresponding to specific components and failure modes based on all online vehicle data, and obtains the actual failure rate of the batch ID based on the batch ID and failure mode corresponding to the specific component. Based on the actual failure rate of the batch ID and the global baseline failure rate corresponding to the specific component and failure mode, the batch quality level of the batch ID corresponding to the specific component is obtained.

[0012] Furthermore, the process of performing a full-dimensional analysis on health-level fault events uploaded to the cloud diagnostic center and generating health-level fault diagnosis results includes: The system pre-determines battery capacity prediction model and motor life prediction model. It inputs the main abnormal events and the five types of related minor abnormalities and operating conditions in the health-level fault events into the battery capacity prediction model and motor life prediction model, respectively. Based on the battery capacity prediction model and motor life prediction model, it outputs the battery capacity decay trend and the remaining life of the motor, respectively, and feeds them back to the vehicle edge node.

[0013] A cloud-edge collaborative new energy vehicle fault diagnosis system includes a cloud-based diagnostic center, which is connected to vehicle-mounted edge nodes and regional edge nodes. The vehicle-mounted edge node is used to collect operating parameters and environmental data of new energy vehicles, and to perform data cleaning, feature extraction and initial fault judgment on the operating parameters. When candidate faults are generated during the initial fault judgment process, all candidate faults are associated and merged to generate one or more fault events. According to the preset static risk value scoring model, the static risk value of each fault event is output, and the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient are obtained. The static risk value is then corrected a second time based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient to obtain the dynamic risk value. The fault level of the fault event is determined based on the dynamic risk value. Based on the preset regional scheduling strategy, the fatal fault events are assigned to the vehicle-mounted edge node for local execution, the warning fault events are assigned to the regional edge node for execution, and the health-level fault events are uploaded to the cloud diagnostic center for execution. The regional edge node is used to perform single-vehicle early warning level fault diagnosis on the received complete data packet, generate fault prediction results, and perform environmental correlation judgment and false alarm suppression based on the fault prediction results; The cloud-based diagnostic center is used to perform batch defect clustering identification on the fault prediction results uploaded by regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the uploaded health-level fault events to generate health-level fault diagnosis results.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. By combining fault event correlation and merging, static risk quantification scoring and multi-dimensional dynamic correction, the comprehensive impact of multi-parameter coupling effects, fault deterioration trends, real-time operating conditions and batch quality differences of components on fault risk is comprehensively considered. This can more realistically restore the actual degree of danger of faults, solve the problem of insufficient adaptability of static thresholds under complex operating conditions, avoid the underestimation of high-risk events caused by multiple fault coupling and thus avoid the false escalation of fault levels caused by instantaneous fluctuations of single parameters, and achieve refined and differentiated accurate classification of fault risks.

[0015] 2. Diagnostic tasks are executed by matching corresponding computing power nodes according to the fault risk level, forming a layered scheduling system where the vehicle-side ensures safety, the regional edge improves accuracy, and the cloud performs in-depth analysis: high-risk faults rely on local closed-loop execution on the vehicle-side to ensure real-time safety response and uphold functional safety standards; medium-risk faults rely on regional edge nodes for in-depth diagnosis and correlation analysis, balancing diagnostic accuracy and response efficiency; long-term health assessments rely on cloud-based big data capabilities for comprehensive analysis. This architecture avoids the problems of insufficient computing power and limited diagnostic accuracy caused by concentrating all computations on the vehicle-side, and also avoids the defects of high transmission latency and strong network dependence caused by uploading all diagnostics to the cloud. While strictly meeting functional safety response requirements, it maximizes the computing power advantages of each layer—cloud and edge—integrating the system.

[0016] 3. Leveraging the multi-vehicle group perception capabilities of regional edge nodes and roadside multi-source monitoring data, it is possible to effectively distinguish between vehicle malfunctions and abnormal signals caused by external factors such as road conditions and weather conditions, eliminating invalid warnings caused by external environmental interference. Simultaneously, through regional clustered screening and full-data statistical analysis in the cloud, common quality risks of components can be identified in the early stages of batch failures, helping automakers to grasp the batch quality status in advance, promptly implement quality intervention, maintenance layout, and risk management, significantly reducing the safety risks and quality losses caused by the spread of batch failures. Attached Figure Description

[0017] Figure 1 This is a flowchart of a cloud-edge collaborative fault diagnosis method for new energy vehicles according to an embodiment of this application; Figure 2 This is a flowchart of a cloud-edge collaborative new energy vehicle fault diagnosis system according to an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] like Figure 1 As shown, the fault diagnosis method for new energy vehicles based on cloud-edge collaboration includes the following steps: Step s1: After the vehicle is powered on, the on-board edge nodes deployed on the new energy vehicle use the on-board CAN bus, Ethernet and dedicated sensor network to synchronously collect the operating parameters and environmental data of the subsystems of the new energy vehicle, such as the battery system, motor system, high voltage electrical system and thermal management system. This includes collecting BMS parameters such as battery cell voltage, temperature, SOC and SOH at a frequency of 10Hz, power system parameters such as motor three-phase current, speed, torque and controller temperature at a frequency of 50Hz, and high voltage safety parameters such as bus voltage ripple and insulation resistance at a frequency of 1kHz. At the same time, it collects contextual data such as ambient temperature and humidity and vehicle operating conditions, and performs data cleaning, feature extraction and initial fault judgment on the operating parameters. Step s2: When candidate faults are generated during the initial fault assessment process, all candidate faults are associated and merged to generate one or more fault events. Based on the preset static risk value scoring model, the static risk value of each fault event is output. Step s3: Obtain the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient. Perform a second correction on the static risk value based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient to obtain the dynamic risk value. Determine the fault level of the fault event based on the dynamic risk value. Step s4: Based on the regional scheduling strategy, critical fault events are assigned to the vehicle edge node for local execution, early warning fault events are assigned to the regional edge node for execution, and health-level fault events are uploaded to the cloud diagnostic center for execution. Step s5: Perform single-vehicle early warning level fault diagnosis on the complete data packets received by the regional edge nodes, generate fault prediction results, and perform environmental correlation discrimination and false alarm suppression operations based on the fault prediction results; Step s6: Perform batch defect clustering identification on the fault prediction results uploaded to the cloud diagnostic center from the regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the health-level fault events uploaded to the cloud diagnostic center to generate health-level fault diagnosis results.

[0020] It should be further explained that, in the specific implementation process, the vehicle edge node is deployed on each new energy vehicle. The hardware carrier is an automotive-grade edge computing domain controller, which adopts an ARM Cortex-A76 multi-core architecture, with a computing power of ≥20 TOPS. It is equipped with 32GB of industrial-grade eMMC storage and a 5G / V2X dual-mode communication module. The vehicle edge node is connected to the vehicle controllers such as BMS, MCU, VCU, and BCM through the CANFD bus, and is connected to sensors such as vehicle cameras and millimeter-wave radar through Ethernet. It is connected to current sensors, vibration sensors, and insulation detection modules through a dedicated analog acquisition channel.

[0021] Regional edge nodes are deployed in locations such as urban roadside units, public charging stations, and highway service areas. The hardware carrier is an industrial-grade edge server with a single node computing power of ≥200 TOPS and a coverage radius of 3~5km. Regional edge nodes communicate with several vehicles in the area through 5G / V2X networks and are connected to the cloud diagnostic center through fiber optic leased lines.

[0022] The cloud-based diagnostic center is deployed on the vehicle manufacturer's private or public cloud platform and adopts a distributed cluster architecture, including a big data storage cluster, a GPU training cluster, a digital twin simulation cluster, and an application service cluster.

[0023] It should be further explained that, in the specific implementation process, the data cleaning and feature extraction process includes: Time base alignment is performed on each type of parameter in the running parameters. After time base alignment, physical range and logical consistency verification, statistical outlier removal, multimodal noise filtering, and hierarchical missing data completion are performed on each type of parameter to obtain the data-cleaned parameters. The time base alignment uses the UTC hard clock of the vehicle domain controller as a unified reference, and marks all data frames with a timestamp with an accuracy of no less than 1ms. Then, based on a 100ms time slice, low-speed data is directly mapped to the corresponding time slice according to the sampling point. High-speed data is segmented according to the time slice, and a complete sampling point sequence is retained in each segment. At the same time, the extreme values ​​and mean values ​​in the time slice are calculated.

[0024] Physical range and logical consistency verification includes: Physical range hard calibration: Set absolute upper and lower limits based on the sensor range and the physical characteristics of the components. Values ​​exceeding the range are directly judged as invalid and discarded. Change rate limit verification: A maximum change rate is set based on the physical characteristics of the component. If the change in adjacent sampling points exceeds the physical limit, it is considered an abnormal sensor jump. For example, a change of >0.3V in battery cell voltage within 100ms or a change of >5℃ in battery temperature within 1s are both impossible jumps and are directly rejected. Logical consistency verification: Based on the vehicle control logic, the matching relationship between parameters is verified, and data with logical contradictions is marked as abnormal. For example: the motor speed is 0 but the output torque is >10Nm, the SOC continues to decrease during charging, the high-voltage contactor is open but there is current in the bus, etc.

[0025] Outlier removal employs a sliding window modified 3σ criterion. An independent sliding statistical window is set for each parameter, with a window length of 10 sampling points for low-speed parameters and 100 sampling points for high-speed parameters. The mean μ and standard deviation σ of the data within the window are calculated in real time, and points exceeding the range of μ±3σ are identified as outliers and removed. For impact-related fault characteristics (such as bearing spalling impact and contact sparking pulse), an impact retention mechanism is added: if outliers with the same characteristics appear in three consecutive windows, they are identified as fault characteristics rather than noise and are retained.

[0026] Multimodal noise filtering includes different filtering algorithms for different types of data based on their noise characteristics: For slowly varying steady-state parameters (temperature, SOC, average unit voltage, etc.): a first-order Kalman filter is used to correct the measured values ​​with the predicted values ​​from the state equation, suppressing random white noise from the sensor while fully preserving the parameter variation trend; For high-speed transient parameters (vibration, ripple, current, etc.): db4 wavelet 3-level decomposition soft thresholding is used for noise reduction, threshold shrinking is performed only on high-frequency detail coefficients, and low-frequency trends and mid-frequency fault impact characteristics are preserved. Switch status quantities (contactors, relays, fault codes, etc.): An anti-jitter filtering mechanism is adopted, and the status change is confirmed only if the status remains consistent for three consecutive sampling cycles, so as to avoid false status jumps caused by bus electromagnetic interference.

[0027] Hierarchical missing data completion includes: Short-term missing data (number of missing points ≤ 3 sampling periods): Nonlinear changes in rapidly changing parameters such as current and voltage are filled by cubic spline interpolation, while slowly changing parameters such as temperature and SOC are filled by linear interpolation. Mid-time missing (3~10 sampling periods): The reduced-order mechanism model at the vehicle edge is used to complete the missing time. For example, when the battery voltage is missing, the output voltage value is estimated by combining the current SOC and charging / discharging current through a first-order RC equivalent circuit model; when the motor temperature is missing, the temperature value is estimated by a loss-thermal resistance model. Long-term missing data (>10 sampling periods): Mark this data segment as a failed segment, exclude it from real-time diagnostic input, and only record the missing event log.

[0028] A dual-scale sliding window is set up. The sliding window for each type of parameter is divided according to the process characteristics of each type of parameter. The dual-scale sliding window includes a fast-changing window and a slow-changing window. The fast-changing window has a window length of 1s and a step size of 0.2s. It corresponds to high-speed sampling data and is used to capture transient faults such as insulation breakdown, power module failure, and bearing impact. The slow-changing window has a window length of 10s and a step size of 2s. It corresponds to low-speed sampling data and is used to capture slow-changing faults such as battery consistency degradation and cooling efficiency decline. Feature extraction was performed on various types of parameters that had undergone data cleaning within different scale windows, resulting in five types of features: time domain features, frequency domain features, time-frequency domain features, degradation trend features, and system coupling correlation features. Among them, time-domain statistical features include basic statistical features (mean, maximum, minimum, peak-to-peak value, root mean square value, standard deviation, variance), impact features (kurtosis, margin factor, impulse factor, peak factor), residual features (mean residual, standard deviation residual, peak residual residual between measured value and output value of reduced-order mechanism model), and distribution features (skewness, coefficient of variation). Frequency domain characteristics: After performing a fast Fourier transform on high-speed data, the fundamental amplitude, the amplitude of the 3rd / 5th / 7th harmonics, the total harmonic distortion rate, and the ratio of the energy of the preset fault characteristic frequency band to the total energy are extracted. The time-frequency domain characteristics are analyzed by using three-layer wavelet packet decomposition to divide the signal into eight frequency bands. The energy value and energy entropy of each frequency band are calculated, with a focus on monitoring energy mutations in the 1~5kHz high-frequency band, corresponding to non-stationary transient faults such as partial discharge of insulation and arcing due to poor contact. The degradation trend characteristics are calculated based on a long window with a length of 1 hour and a step size of 10 minutes, including the linear change rate of parameters, cumulative offset, and degradation acceleration, which are used to characterize the long-term performance degradation trend of the component. System coupling correlation characteristics, including Pearson correlation coefficients of cross-system parameters and mapping residuals of measured points deviating from the baseline mapping curve under normal operating conditions, are used to identify multi-system coupling faults.

[0029] It should be further explained that, in the specific implementation process, the initial fault diagnosis process includes: The system includes a pre-defined hard threshold rule base, a dynamic threshold rule base, and a health-level rule base. The hard threshold rule base is constructed based on national safety standards and vehicle safety design limits. It represents an absolute safety red line and is immediately marked as a fatal candidate upon triggering, with the highest priority. The system compares the temporal statistical characteristics of each type of parameter with the hard threshold corresponding to the temporal statistical characteristics of each type of parameter in the hard threshold rule base. When the temporal statistical characteristics of a certain type of parameter reach the hard threshold within two consecutive corresponding scale sliding windows, that type of parameter is marked as a fatal candidate fault. The dynamic threshold library includes warning thresholds corresponding to the time-domain statistical characteristics of various types of parameters under different operating conditions. The operating conditions of new energy vehicles are obtained based on environmental data. The operating conditions include vehicle driving status (charging / idling / constant speed / acceleration / deceleration), ambient temperature (temperature range, altitude, humidity), and component status (SOH range, cumulative mileage, charge / discharge rate). The warning thresholds in the dynamic threshold library are called according to the current operating conditions. The time-domain statistical characteristics of each type of parameter are compared with the warning thresholds corresponding to the time-domain statistical characteristics of each type of parameter. When the time-domain statistical characteristics of a certain type of parameter reach the warning threshold within two consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the warning level. The degradation trend characteristics of each type of parameter (such as the battery SOH decay rate) are compared with the thresholds corresponding to the degradation trend characteristics of each type of parameter in the health level rule base. When the degradation trend characteristic of a certain type of parameter reaches the threshold within three consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the health level.

[0030] The system can quickly filter out more than 95% of normal data from tens of thousands of samples per second, leaving only candidate faults, thus solving the problem of "large data volume and limited computing power". The core of this technology is speed.

[0031] It should be further explained that, in the specific implementation process, when candidate faults are generated during the initial fault assessment, the process of associating and merging all candidate faults to generate one or more fault events, and outputting the static risk value of each fault event according to the preset static risk value scoring model includes: The core objective of initial fault diagnosis is to pick out abnormal samples from massive amounts of normal data. It only addresses the question of whether there are abnormalities, not how dangerous the abnormalities are. Any parameter exceeding the limit is marked as a candidate. For example, the initial judgment outputs of "50mV differential voltage between individual battery cells" and "200mV differential voltage", which belong to the same warning level, are indistinguishable. They cannot distinguish between slight degradation and critical faults that are about to deteriorate. Furthermore, the initial judgment is triggered by individual parameters and does not automatically assess the risk superposition effect caused by the simultaneous abnormality of multiple parameters such as "temperature rise + voltage drop + accelerated temperature rise".

[0032] When any candidate fault (including fatal / warning / health) is generated, a merging time window with a total duration of 1.5 seconds is constructed, centered on the generation timestamp of the candidate fault, taking 0.5 seconds forward and 1 second backward. This window length covers the maximum time delay of fault propagation in new energy vehicles (such as the time difference of heat transfer and electrical coupling), and can completely accommodate multiple parameter anomalies triggered successively by the same fault. All candidate faults generated within the merging time window, regardless of their initial judgment level or subsystem, are merged into the merging candidate pool. Based on the preset vehicle fault domain classification system, the candidate faults in the merging candidate pool are clustered to generate one or more fault events. Specifically, for multiple candidate faults triggered by multiple parameter anomalies within the same sliding time window, according to the vehicle fault domain classification system (battery system, motor system, high voltage system, thermal management system), all candidate faults under the same component are forcibly classified into the same fault event. For example, single cell temperature exceeding the limit (warning level), temperature rise rate exceeding the standard (fatal level), and single cell voltage drop (warning level) all belong to the "battery system - single cell" component and are automatically classified into the same fault event. Based on the initial judgment level of each candidate fault in each fault event, determine the primary anomaly (only one) and associated secondary anomalies in each fault event. The priority determination rules for the primary anomaly are as follows (from high to low): 1. Priority is given to the initial judgment level: Fatal level candidates > Warning level candidates > Healthy level candidates; 2. Priority is given to the degree of deviation under the same initial judgment level: the greater the proportion of feature parameters exceeding the threshold, the higher the priority; the scoring priority of each fault event is determined based on the initial judgment level of the primary anomaly in each fault event (fatal level candidates > warning level candidates, and priority is given to the degree of deviation under the same initial judgment level). A static risk value scoring model is pre-installed. This model is a static knowledge base on the vehicle edge node and serves as the anchor benchmark for scoring. It is uniformly defined and synchronously updated by the cloud diagnostic center based on global fault data and includes the safety benchmark scores corresponding to all known fault events. According to the scoring priority, starting with the highest priority, the main anomaly event and the associated secondary anomaly data packets of each fault event are sequentially input into the static risk value scoring model. The static risk value of each fault event is output based on the static risk value scoring model. The data packet includes time-domain statistical characteristics, degradation trend characteristics, operating conditions, and preliminary assessment level.

[0033] The static risk value scoring model operates as follows: For each primary anomaly and associated secondary anomaly within a fault event, calculate the security score for each primary anomaly and associated secondary anomaly separately: ,in: : The security score of the main exception event / related minor exception; The safety baseline score for the main abnormal event / related minor abnormality (score corresponding to moderate severity, such as 60 points); The main abnormal event / related event is scored from the upper limit of the abnormality's security level (i.e., the score when it completely fails, such as 100 points). : Measured values ​​of current feature parameters; The threshold that triggers the initial judgment; : Safety limit threshold of feature parameters.

[0034] Based on the security score corresponding to each primary anomaly and associated secondary anomaly, a weighted average formula of primary anomaly + associated secondary anomaly is used to calculate the final score, with a maximum score of 100 points. ,in: Static risk value of the failure event: : Safety score of the main abnormal event; The i-th association is derived from the security score of the anomaly; : The fusion weight of the i-th associated anomaly; the total score is capped at 100 points to avoid infinite accumulation; the fusion weight of associated anomalies is set according to the initial judgment level of the associated anomaly: the ratio of the corresponding fusion weights for associated anomalies with the initial judgment levels of fatal, warning, and healthy is 3:2:1, and the total fusion weight of all associated anomalies does not exceed 0.5.

[0035] It should be further explained that, in the specific implementation process, the process of obtaining the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient, and then performing a secondary correction on the static risk value based on these coefficients to obtain the dynamic risk value, and finally determining the fault level of the fault event based on the dynamic risk value, includes: A pre-defined fault parameter coupling correlation graph is used, with the core monitoring parameters of each subsystem of the vehicle as the smallest nodes. It covers the core parameters of the four major systems: battery, motor, high-voltage electrical, and thermal management (such as cell voltage, cell temperature, bus ripple, motor vibration acceleration, contactor conduction voltage drop, etc.). The directed edges of the fault parameter coupling correlation graph represent the fault propagation and causal coupling relationship between parameters. Each edge has three core attributes: 1. Coupling strength: divided into three levels: strong coupling, medium coupling, and weak coupling, corresponding to the inevitability and degree of impact of fault propagation; 2. Propagation direction: clearly defines the "cause → effect" propagation sequence (such as "increased contact resistance → increased bus ripple", which is irreversible); 3. Typical propagation delay: the time constant (milliseconds to minutes) for fault propagation from upstream parameters to downstream parameters.

[0036] Using the type parameters corresponding to the main anomaly and associated secondary anomalies in the fault event as nodes, a connection path between nodes is searched in the fault parameter coupling association graph. The coupling strength between nodes is determined based on the path length and edge attributes, and a coupling correction coefficient is obtained based on the coupling strength between nodes. ; Specifically, if there is a direct directed edge (path length = 1) between any two abnormal parameters and the edge attribute is strong coupling, it is determined to be strong coupling.

[0037] If there is an indirect transmission path between any two abnormal parameters (path length = 2~3), or if there is a direct directed edge between any two abnormal parameters and the edge attribute is medium coupling, it is determined to be medium coupling.

[0038] If there is no direct / indirect transmission link between any two abnormal parameters, or if there is a direct directed edge between any two abnormal parameters and the edge attribute is weak coupling, it is determined to be weak coupling.

[0039] Based on the coupling strength between abnormal parameters, map the corresponding coupling correction coefficients. If the coupling strength is strong coupling, then =1.3, if the coupling strength is medium coupling, then =1.15, if the coupling strength is weak, then =1.0; It should be further noted that for scenarios with three or more abnormal parameters, the coupling strength of the longest coupled link is used as the standard, with an additional link length correction: for each additional coupled node, the coefficient increases by 0.05, up to a maximum of 1.5. For example, three strongly coupled nodes connected in series... .

[0040] Historical temporal statistical features of the type parameters corresponding to the main abnormal events within the first k (k=30) fast-changing windows of the merged time window are extracted. Operating condition normalization is performed on the historical temporal statistical features within each fast-changing window to obtain the relative degradation degree of each fast-changing window. Specifically, the vehicle edge node has a built-in operating condition-feature baseline mapping table, which pre-stores normal feature baseline values ​​under different operating conditions. The current operating condition is read in real time, and the baseline value of the temporal statistical features under the corresponding operating condition is obtained by looking up the table. Calculate the relative degradation of the current window. : ,in These are time-domain statistical characteristics. This indicates a deviation from the normal baseline; the larger the value, the more severe the degradation. This indicates performance better than the baseline. A linear fit is performed on the relative degradation degree of each rapid-change window to obtain the degradation rate of the main anomaly event. Specifically, the rapid-change window number is used as the independent variable. The relative degradation of the fast-changing window is the dependent variable. The linear trend equation was obtained by fitting: ,in: The slope of the degradation rate, in % / s; The degradation continues to increase, and the failure is showing a worsening trend; The degradation level is basically stable, with no obvious trend. The degradation rate continues to decrease, and the failure is showing a mitigating trend. The value is the intercept and has no direct physical meaning. The degradation rate is normalized, and the degradation trend correction coefficient is obtained based on the normalized degradation rate. ; For each type of fault, a preset baseline degradation rate is established. (This value is obtained by the cloud-based diagnostic center based on massive amounts of fault data and is simultaneously distributed to the edge as the model is updated.) It represents the average rate at which this type of fault progresses from the warning level to the critical level, and the rate of degradation. The absolute value and the baseline degradation rate By comparison, the normalized degradation rate multiplier is obtained. : , The larger the value, the higher the current rate of degradation, which far exceeds the normal aging level, and the higher the risk. Deterioration trend correction coefficient The mapping rule is: when k is greater than 0, if <0.1 (almost no degradation), then =1.0; if If it falls within the range of 0.1 to 0.5, then =1.1; if If it falls within the range of 0.5 to 1.0 (excluding 0.5), then =1.2; if >1.0, then =1.3; When k is less than 0, if <0.1 (almost no degradation), then =1.0; if If it falls within the range of 0.1 to 0.5, then =0.95; if If it falls within the range of 0.5 to 1.0 (excluding 0.5), then =0.9; if >1.0, then =0.85.

[0041] It should be noted that the preset parameters and preset thresholds in this invention are set by those skilled in the art based on actual conditions or obtained through simulation with a large amount of data.

[0042] A preset working condition correction factor mapping table is provided, which includes working condition correction factors corresponding to different working conditions. The working condition correction factor matching the current working condition is obtained based on the working condition correction factor mapping table. ; Batch quality correction factor It is the only coefficient in the dynamic fault risk correction system that relies on cross-vehicle group statistical data. Its core function is to incorporate batch-specific quality defects of parts into the real-time fault risk assessment of a single vehicle, solving the industry pain point that a single vehicle cannot perceive common quality risks in the same batch, and realizing the joint rating of single vehicle abnormal characteristics + batch quality baseline.

[0043] Obtain the batch ID of the specific component corresponding to the main anomaly event in the fault event; retrieve the batch quality level corresponding to the batch ID of the specific component from the cloud diagnostic center; and obtain the batch quality correction coefficient based on the batch quality level. ; If the batch quality grade is normal, then the batch quality correction factor is... =1.0; If the batch quality level is a batch of concern, then the batch quality correction factor is 1.0. =1.05; If the batch quality level is a warning batch, then the batch quality correction factor is 1.05. =1.1; If the batch quality level is high-risk, then the batch quality correction factor is 1.1. =1.2; Based on the coupling correction coefficient Deterioration trend correction coefficient Operating condition correction factor and batch quality correction factor The static risk value is corrected to obtain the dynamic risk value; ;in, This is the static risk value. This is a dynamic risk value.

[0044] Within the dynamic risk value range (0-100 points), threshold points are selected to divide the fault into sub-intervals of different fault levels. The fault level of the fault event is determined according to the sub-interval to which the dynamic risk value belongs. The fault levels include fatal level (dynamic risk value ≥ 80 points), warning level (40 points ≤ dynamic risk value < 80 points), and healthy level (dynamic risk value < 40 points).

[0045] It should be further explained that, in the specific implementation process, based on the regional scheduling strategy, the process of allocating fatal fault events to the vehicle edge node for local execution, early warning fault events to the regional edge node for execution, and health-level fault events to the cloud diagnostic center for execution includes: For catastrophic fault events, the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the catastrophic fault event are input into a local fault diagnosis model deployed on the vehicle edge node. The local fault diagnosis model outputs fault judgment results (including fault type, fault confidence and fault severity level). Based on the fault judgment results, protection control commands are generated. Specifically, the vehicle edge node matches and generates protection control commands of the corresponding level according to a preset catastrophic fault-protection control command mapping table (existing technology) and sends them to execution units such as the vehicle controller (VCU), battery management system (BMS), and motor controller (MCU). The fault event summary of the fault judgment results is extracted and sent to the regional edge node and cloud diagnostic center currently connected to the new energy vehicle. For critical faults, a "local closed-loop + simplified reporting" allocation mode is adopted. The entire process of fault type identification, severity determination, emergency protection command generation and issuance is executed entirely by a dedicated isolated computing core for diagnosis deployed on the vehicle edge node. This computing core is physically isolated from functions such as autonomous driving and smart cockpit at the hardware level, and its exclusive resource allocation is not affected by the load of other systems, ensuring that the end-to-end latency from fault feature triggering to protection command output to the vehicle controller is ≤10ms. After diagnosis, the vehicle edge node only uploads a simplified, anonymized fault event summary to the regional edge node and the cloud diagnostic center. Each message includes a timestamp, fault type, confidence level, and coarse-grained location information, with a data size of less than 1KB. Furthermore, this strategy sets rigid safety constraints: regardless of subsequent network fluctuations, the diagnostic and protection execution authority for critical faults must not be transferred to the regional node or the cloud. The vehicle must always retain independent closed-loop capabilities as the ultimate bottom line for functional safety.

[0046] The local fault diagnosis model used in this invention belongs to the conventional and mature technology in the field of vehicle fault diagnosis. Specifically, it is a lightweight automotive-grade diagnostic model optimized for fatal fault scenarios. The overall model adopts a two-level cascaded architecture of "rule pre-verification + lightweight temporal network inference". After INT8 quantization compression and functional safety verification, it is deployed on the diagnostic dedicated isolated computing core of the vehicle edge node. The single sample inference latency is ≤5ms, which meets the timing requirements of millisecond-level response for fatal faults and meets the ISO26262ASIL-B level requirements.

[0047] For warning-level fault events, the main abnormal event and the associated complete data packet of the warning-level fault event are uploaded to the edge node of the area currently connected to the new energy vehicle. The complete data packet includes five types of features, operating conditions, vehicle desensitization identifier and GPS location stamp. For health-level fault events, based on the operating conditions of new energy vehicles, the main abnormal events and associated five types of abnormal characteristics and operating conditions in the health-level fault events are uploaded to the cloud diagnostic center. Specifically, complete data packets are uploaded in batches every day when the vehicle is charging or parked, without occupying the communication bandwidth during driving.

[0048] Among them, the vehicle desensitization label is a secondary desensitization process that performs on the vehicle identification information, retaining only statistical attributes such as vehicle model, component batch corresponding to the fault event, cumulative mileage, and operating area, while removing sensitive information that can identify the individual vehicle, thus meeting the requirements for automotive data security compliance.

[0049] It should be further explained that, in the specific implementation process, the process of performing single-vehicle early warning-level fault diagnosis on the complete data packets received by the regional edge nodes, generating fault prediction results, and performing environmental correlation discrimination and false alarm suppression operations based on the fault prediction results includes: When the edge node receives the complete data packet, it extracts the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the data packet, as well as the associated anomalies. It generates a multi-dimensional joint feature vector and inputs the multi-dimensional joint feature vector into the multi-scale attention diagnosis model (existing mature technology) deployed on the edge node. The multi-scale attention diagnosis model outputs the fault prediction result, which includes specific components and fault modes. The fault prediction result is then combined with the vehicle desensitization identifier, GPS location stamp and operating condition in the complete data packet to bind the vehicle desensitization identifier, GPS location stamp and operating condition. The fault prediction results corresponding to each vehicle's desensitized identifier are screened for fault morphology (the fault morphology must be transient single impact type). If the fault prediction result corresponding to the vehicle's desensitized identifier does not meet the fault morphology requirements (not transient single impact type), then temporary handling suggestions are generated based on the fault prediction results corresponding to the vehicle's desensitized identifier and the operating conditions. Specifically, after completing the root cause localization and severity assessment of the warning-level fault, the regional edge node, combined with the vehicle's current operating conditions, matches the preset fault-operating condition-handling mapping rules (existing technology) to generate executable temporary handling suggestions for drivers, passengers, and the vehicle controller; and feeds them back to the corresponding vehicle edge node (the vehicle edge node to be fed back is determined based on the vehicle desensitized identifier bound to the fault prediction result). If the fault prediction result meets the fault morphology requirements, then the following steps are executed: Step 1: Spatially divide the coverage area of ​​the regional edge node into several road segment units. Specifically, continuous roads are divided into fixed-length road segment units according to road station numbers, with a default segment length of 500m. The units are also differentiated by travel direction (up or down), and each segment unit is assigned a unique segment ID. For special road segments such as intersections, ramps, and construction sections, they are further subdivided into 200m short units. Vehicle GPS locations are mapped to the corresponding road segment ID and lane number after map matching, with a location matching accuracy ≤10m. The complete data packets uploaded to the regional edge node are divided temporally, into different time windows using a sliding time window mechanism. The default window length is 5 minutes, and the sliding step is 1 minute. The complete data packets within each road segment unit are continuously and continuously analyzed. The fault prediction results corresponding to the complete data packet are bound to the time window of that complete data packet. Step 2: Cluster the fault prediction results corresponding to the de-identified identifiers of each vehicle within the same time window and the same road segment unit, group the same fault prediction results into the same group, and obtain the number of fault prediction results in each group. The unique identifier of the de-identified vehicle is used as the statistical unit. If the same vehicle generates the same type of fault multiple times within the same time and space window, it is only counted as 1 valid event. If the number of fault prediction results in a certain group is greater than the preset warning threshold for the same type, then proceed to Step 3. If the number of fault prediction results in each group is less than or equal to the preset warning threshold for the same type, then proceed to Step 5. The warning thresholds for the same type are dynamically set based on the road type of each road segment unit: highways and expressways: default threshold of 4 units; urban arterial roads and ordinary roads: default threshold of 3 units; construction sections and sections with dense speed bumps: threshold increased to 5 units.

[0050] Step 3: Acquire roadside monitoring data for a specific road segment unit. Generate an environmental causation quantitative score based on the roadside monitoring data and fault prediction results. The roadside monitoring data includes road surface smoothness index (directly measured by road surface embedded sensors, RSU vehicle-mounted high-definition cameras + AI recognition modules, etc.), meteorological environmental data (instantaneous wind speed and direction, rainfall, ambient temperature, air humidity, visibility, etc., obtained through regional meteorological stations), and traffic condition data (including average vehicle speed, traffic volume, congestion level, etc., collected through roadside traffic radar and checkpoint detection equipment). Subsequently, based on the fault prediction results, a linear scoring framework is adopted, employing independent scoring for each item and weighted summation. The inputs are roadside monitoring data and fault prediction results, and the output is a total environmental causation score ranging from 0 to 100. A higher score indicates a greater likelihood that the current anomaly is caused by environmental factors. The specific execution process is as follows: For different types of susceptible faults, corresponding scoring dimensions and weight allocation tables are pre-established, with the sum of the weights of all dimensions being 100%. The dimension division and weight values ​​are determined by those skilled in the art based on the fault mechanism, the degree of environmental impact, and engineering experience, and are considered routine engineering parameter configurations.

[0051] Taking suspension / chassis impact-related faults as an example, the scoring dimensions are divided into four categories: road surface condition, road facilities, vehicle speed, and road surface slippage, with corresponding weights of 40%, 20%, 20%, and 20%, respectively. For each specific monitoring indicator under each scoring dimension, normalization is performed based on preset segmented thresholds, mapping the original monitoring values ​​of different dimensions to sub-scores ranging from 0 to 100. The more an indicator deviates from normal operating conditions, the stronger its interference with fault characteristics, and the higher the corresponding score. For example: For road surface smoothness index: Based on the International Roughness Index (IRI), multiple segmented thresholds are set. The smoother the road surface, the lower the score, and the higher the bumpiness, the higher the score. The full score corresponds to severe potholes / damage. For the vehicle speed index: the score is divided into segments according to the vehicle speed range. The higher the vehicle speed, the greater the amplitude of the road impact transmission, and the higher the corresponding score. The impact is small under low speed conditions, and the corresponding score is low. For meteorological indicators such as rainfall and wind speed: scores are awarded in segments according to meteorological levels. The more severe the meteorological conditions, the stronger the interference with vehicle operation and sensor signals, and the higher the corresponding score.

[0052] The final environmental factor score is obtained by multiplying the scores of each dimension by their corresponding weights and then summing them. The calculation formula is as follows: ,in, The total score is based on environmental factors. Let be the weight coefficient of the i-th rating dimension. Let be the score for the i-th dimension, and n be the total number of scoring dimensions. The calculated total score ranges from 0 to 100.

[0053] The above-mentioned segmentation thresholds and weighting coefficients are all general empirical values ​​in the fields of vehicle engineering and road engineering. They can be adaptively adjusted according to vehicle type and road grade without requiring creative effort.

[0054] If the environmental causation quantification score is greater than the preset environmental causation quantification threshold, then a multi-vehicle coverage determination is performed on a certain group. Specifically, the batch ID corresponding to the vehicle model and the specific component is determined according to the vehicle desensitization label. If the vehicles in a group cover ≥2 vehicle models and ≥2 different batch IDs, then no batch clustering is determined. If there is only a single vehicle model and a single specific component batch ID in a group, then batch clustering is determined. If a certain group is determined to have no batch clustering, then step four is executed. If a certain group is determined to have batch clustering, then the fault prediction result of a certain group is uploaded to the cloud diagnostic center to perform batch defect clustering identification. If the environmental causation quantification score is less than or equal to the preset environmental causation quantification threshold, then step five is executed. Step 4: Remove the fault prediction results corresponding to the desensitized labels of each vehicle in a certain group and mark them as environmental interference; Step 5: For each vehicle desensitized label in each group, based on the fault prediction results and operating conditions corresponding to the vehicle desensitized label, generate temporary handling suggestions for the vehicle desensitized label and feed them back to the corresponding vehicle edge node.

[0055] It should be further explained that, in the specific implementation process, the process of performing batch defect clustering identification on the fault prediction results uploaded by the regional edge nodes and generating batch quality levels corresponding to different batch IDs includes: Based on the fault prediction results corresponding to the vehicle desensitization labels in a certain group, the specific components and fault modes are determined. The batch IDs corresponding to the specific components are extracted from the vehicle desensitization labels. The specific components, the batch IDs corresponding to the specific components, and the fault modes are uploaded to the cloud diagnostic center. The cloud-based diagnostic center uses data from all vehicles on the network to obtain a global baseline failure rate corresponding to specific components and failure modes. Specifically, it calculates the global baseline failure rate by rolling over the past 12 months of operating data, excluding extreme data from new vehicle break-in periods and older vehicles exceeding their service life. : And based on the batch ID and failure mode corresponding to the specific component, the actual failure rate of the batch ID is obtained. Specifically: Based on the actual failure rate of the batch ID and the global baseline failure rate corresponding to the specific component and failure mode, obtain the batch quality level of the batch ID corresponding to the specific component.

[0056] The batch quality grade mapping rule is: if If the batch quality grade is normal, then the batch quality grade is normal. Then the batch quality level is a batch of concern; if If the batch quality level is a warning batch; If so, the batch quality level is high-risk batch.

[0057] Upgrading from independent vehicle diagnostics to collaborative group perception reduces false alarms caused by environmental factors and shortens the discovery cycle of batch quality defects from months to hours, significantly reducing recall costs and brand risks for automakers.

[0058] It should be further explained that, in the specific implementation process, the process of performing a full-dimensional analysis of health-level fault events uploaded to the cloud diagnostic center and generating health-level fault diagnosis results includes: This stage is the final execution stage of the health-level fault task, deployed in the cloud diagnostic center. Relying on the full lifecycle operation data accumulated by the cloud-edge collaborative architecture, it conducts long-term health assessments and lifespan predictions. The battery capacity degradation prediction and motor remaining lifespan prediction models involved are existing mature technologies in the field of new energy vehicle health management. The innovation of this invention does not lie in the prediction model itself, but in providing high-quality standardized data input for cloud prediction with extremely low bandwidth overhead through edge feature extraction and idle-time batch upload mechanism, and realizing reverse closed-loop control of prediction results to the edge.

[0059] The system pre-determines battery capacity prediction model and motor life prediction model. It inputs the main abnormal events and the five types of related minor abnormalities and operating conditions in the health-level fault events into the battery capacity prediction model and motor life prediction model, respectively. Based on the battery capacity prediction model and motor life prediction model, it outputs the battery capacity decay trend and the remaining life of the motor, respectively, and feeds them back to the vehicle edge node.

[0060] like Figure 2 As shown, the new energy vehicle fault diagnosis system based on cloud-edge collaboration includes a cloud-based diagnostic center, which is connected to vehicle-mounted edge nodes and regional edge nodes. The vehicle-mounted edge node is used to collect operating parameters and environmental data of new energy vehicles, and to perform data cleaning, feature extraction and initial fault judgment on the operating parameters. When candidate faults are generated during the initial fault judgment process, all candidate faults are associated and merged to generate one or more fault events. According to the preset static risk value scoring model, the static risk value of each fault event is output, and the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient are obtained. The static risk value is then corrected a second time based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient to obtain the dynamic risk value. The fault level of the fault event is determined based on the dynamic risk value. Based on the preset regional scheduling strategy, the fatal fault events are assigned to the vehicle-mounted edge node for local execution, the warning fault events are assigned to the regional edge node for execution, and the health-level fault events are uploaded to the cloud diagnostic center for execution. The regional edge node is used to perform single-vehicle early warning level fault diagnosis on the received complete data packet, generate fault prediction results, and perform environmental correlation judgment and false alarm suppression based on the fault prediction results; The cloud-based diagnostic center is used to perform batch defect clustering identification on the fault prediction results uploaded by regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the uploaded health-level fault events to generate health-level fault diagnosis results.

[0061] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A new energy vehicle fault diagnosis method based on cloud edge collaboration, characterized in that, Includes the following steps: Step s1: Collect the operating parameters and environmental data of new energy vehicles by deploying on-board edge nodes on the new energy vehicles, and perform data cleaning, feature extraction and preliminary fault judgment on the operating parameters; Step s2: When candidate faults are generated during the initial fault assessment process, all candidate faults are associated and merged to generate one or more fault events. Based on the preset static risk value scoring model, the static risk value of each fault event is output. Step s3: Obtain the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient. Perform a second correction on the static risk value based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient, and batch quality correction coefficient to obtain the dynamic risk value. Determine the fault level of the fault event based on the dynamic risk value. Step s4: Based on the regional scheduling strategy, critical fault events are assigned to the vehicle edge node for local execution, early warning fault events are assigned to the regional edge node for execution, and health-level fault events are uploaded to the cloud diagnostic center for execution. Step s5: Perform single-vehicle early warning level fault diagnosis on the complete data packets received by the regional edge nodes, generate fault prediction results, and perform environmental correlation discrimination and false alarm suppression operations based on the fault prediction results; Step s6: Perform batch defect clustering identification on the fault prediction results uploaded to the cloud diagnostic center from the regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the health-level fault events uploaded to the cloud diagnostic center to generate health-level fault diagnosis results.

2. The cloud-edge collaboration based new energy vehicle fault diagnosis method according to claim 1, characterized in that, The process of data cleaning and feature extraction includes: Time base alignment is performed on each type of parameter in the running parameters. After time base alignment, physical range and logical consistency verification, statistical outlier removal, multimodal noise filtering, and hierarchical missing data completion are performed on each type of parameter to obtain the data-cleaned parameters. A dual-scale sliding window is set up, and the scale sliding window corresponding to each type of parameter is divided according to the process characteristics of each type of parameter. The dual-scale sliding window includes a fast-changing window and a slow-changing window. Feature extraction is performed on each type of parameter that has completed data cleaning within different scale windows to obtain five types of features, including time domain features, frequency domain features, time-frequency domain features, degradation trend features, and system coupling correlation features.

3. The cloud-edge collaboration based new energy vehicle fault diagnosis method according to claim 2, characterized in that, The process of initial fault diagnosis includes: The system includes a pre-defined hard threshold rule base, a dynamic threshold rule base, and a health-level rule base. It compares the temporal statistical characteristics of each type of parameter with the hard threshold corresponding to the temporal statistical characteristics of each type of parameter in the hard threshold rule base. When the temporal statistical characteristics of a certain type of parameter reach the hard threshold within two consecutive corresponding scale sliding windows, the certain type of parameter is marked as a fatal candidate fault. The dynamic threshold library includes warning thresholds corresponding to the time-domain statistical characteristics of various types of parameters under different operating conditions. The operating conditions of new energy vehicles are obtained based on environmental data. The warning thresholds in the dynamic threshold library are called according to the current operating conditions. The time-domain statistical characteristics of each type of parameter are compared with the warning thresholds corresponding to the time-domain statistical characteristics of each type of parameter. When the time-domain statistical characteristics of a certain type of parameter reach the warning threshold within two consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the warning level. The degradation trend characteristics of each type of parameter are compared with the thresholds corresponding to the degradation trend characteristics of each type of parameter in the health-level rule base. When the degradation trend characteristics of a certain type of parameter reach the threshold within three consecutive corresponding scale sliding windows, the certain type of parameter is marked as a candidate fault for the health level.

4. The cloud-edge collaboration based new energy vehicle fault diagnosis method according to claim 3, characterized in that, When candidate faults are generated during the initial fault assessment process, the process of associating and merging all candidate faults to generate one or more fault events, and outputting the static risk value of each fault event according to the preset static risk value scoring model includes: When any candidate fault is generated, a merging time window is constructed with the generation timestamp of the candidate fault as the center. All generated candidate faults within the merging time window are merged into the merging candidate pool. Based on the preset vehicle fault domain classification system, the candidate faults in the merging candidate pool are clustered to generate one or more fault events. The primary anomaly and associated secondary anomalies in each fault event are determined based on the initial judgment level of each candidate fault in each fault event, and the scoring priority of each fault event is determined based on the initial judgment level of the primary anomaly in each fault event. A static risk value scoring model is preset. According to the scoring priority, the main abnormal event and the associated secondary abnormal data packets in each fault event are input into the static risk value scoring model. The static risk value of each fault event is output according to the static risk value scoring model. The data packet includes time-domain statistical characteristics, degradation trend characteristics, operating conditions, and preliminary assessment level.

5. The cloud-edge collaboration based new energy vehicle fault diagnosis method according to claim 4, characterized in that, The process of obtaining coupling correction coefficients, degradation trend correction coefficients, operating condition correction coefficients, and batch quality correction coefficients; performing secondary correction on the static risk value based on these coefficients to obtain the dynamic risk value; and determining the fault level of the fault event based on the dynamic risk value includes: A fault parameter coupling association graph is preset. The type parameters corresponding to the main abnormal event and the associated secondary abnormal event in the fault event are used as nodes. The connection path between nodes is searched in the fault parameter coupling association graph. The coupling strength between nodes is determined according to the path length and edge attributes. The coupling correction coefficient is obtained according to the coupling strength between nodes. Extract the historical time-domain statistical features of the type parameters corresponding to the main abnormal events within the first k fast-changing windows of the merged time window, perform working condition normalization on the historical time-domain statistical features within each fast-changing window, obtain the relative degradation degree of each fast-changing window, perform linear fitting on the relative degradation degree of each fast-changing window, obtain the degradation rate of the main abnormal event, perform normalization on the degradation rate, and obtain the degradation trend correction coefficient based on the degradation rate after normalization. A preset working condition correction coefficient mapping table is provided, which includes working condition correction coefficients corresponding to different working conditions. The working condition correction coefficient matching the current working condition is obtained based on the working condition correction coefficient mapping table. Obtain the batch ID of the specific component corresponding to the main abnormal event in the fault event, obtain the batch quality level corresponding to the batch ID of the specific component from the cloud diagnostic center, and obtain the batch quality correction coefficient based on the batch quality level. The static risk value is corrected based on the coupling correction coefficient, the degradation trend correction coefficient, the operating condition correction coefficient, and the batch quality correction coefficient to obtain the dynamic risk value. Within the dynamic risk value range, threshold points are selected to divide the fault into sub-intervals of different fault levels. The fault level of the fault event is determined according to the sub-interval to which the dynamic risk value belongs. The fault levels include fatal, warning, and healthy levels.

6. The new energy vehicle fault diagnosis method based on cloud-edge collaboration according to claim 5, characterized in that, Based on the regional scheduling strategy, the process of allocating critical fault events to onboard edge nodes for local execution, early warning fault events to regional edge nodes for execution, and health-level fault events to the cloud diagnostic center for execution includes: For critical fault events, the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the critical fault event are input into the local fault diagnosis model deployed on the vehicle edge node. The fault judgment result is output according to the local fault diagnosis model. The protection control command is generated according to the fault judgment result, and the fault event summary of the fault judgment result is extracted and sent to the regional edge node and cloud diagnosis center currently connected to the new energy vehicle. For warning-level fault events, the main abnormal event and the associated complete data packet of the warning-level fault event are uploaded to the edge node of the area currently connected to the new energy vehicle. The complete data packet includes five types of features, operating conditions, vehicle desensitization identifier and GPS location stamp. For health-level fault events, based on the operating conditions of new energy vehicles, the main abnormal events and associated five types of abnormal characteristics and operating conditions in the health-level fault events are uploaded to the cloud diagnostic center.

7. The new energy vehicle fault diagnosis method based on cloud-edge collaboration according to claim 6, characterized in that, The process of performing single-vehicle early warning-level fault diagnosis on complete data packets received by regional edge nodes, generating fault prediction results, and performing environmental correlation discrimination and false alarm suppression based on the fault prediction results includes: When the edge node receives the complete data packet, it extracts the time-domain features, frequency-domain features, time-frequency-domain features, system coupling correlation features and degradation trend features of the main abnormal event in the data packet, as well as the associated anomalies. It generates a multi-dimensional joint feature vector and inputs the multi-dimensional joint feature vector into the multi-scale attention diagnosis model deployed on the edge node. The multi-scale attention diagnosis model outputs the fault prediction result, which includes specific components and fault modes. The fault prediction result is then bound to the vehicle desensitization identifier, GPS location stamp and operating condition in the complete data packet. The fault prediction results corresponding to each vehicle's desensitized identifier are screened for fault modes. If the fault prediction result corresponding to the vehicle's desensitized identifier does not meet the fault mode requirements, a temporary handling suggestion is generated based on the fault prediction result corresponding to the vehicle's desensitized identifier and the operating conditions, and fed back to the corresponding vehicle edge node. If the fault prediction result meets the fault mode requirements, the following steps are executed: Step 1: Spatially divide the coverage area of ​​the regional edge node into several road segment units. Divide the complete data packets uploaded to the regional edge node into time-divided data packets in different time windows. Bind the fault prediction results corresponding to the complete data packets to the time window according to the time window to which the complete data packets belong. Step 2: Cluster the fault prediction results corresponding to the desensitized labels of each vehicle within the same time window and the same road segment unit, group the same fault prediction results into the same group, and obtain the number of fault prediction results in each group. If the number of fault prediction results in a certain group is greater than the preset warning threshold for the same type, then proceed to Step 3. If the number of fault prediction results in each group is less than or equal to the preset warning threshold for the same type, then proceed to Step 5. Step 3: Obtain roadside monitoring data for a specific road segment unit. Generate an environmental cause quantitative score based on the roadside monitoring data and fault prediction results. If the environmental cause quantitative score is greater than the preset environmental cause quantitative threshold, perform multi-vehicle coverage determination for the specific group. If the specific group is determined to have no batch clustering, proceed to Step 4. If the specific group is determined to have batch clustering, upload the fault prediction results of the specific group to the cloud diagnostic center to perform batch defect clustering identification. If the environmental cause quantitative score is less than or equal to the preset environmental cause quantitative threshold, proceed to Step 5. Step 4: Remove the fault prediction results corresponding to the desensitized labels of each vehicle in a certain group; Step 5: For each vehicle desensitized label in each group, based on the fault prediction results and operating conditions corresponding to the vehicle desensitized label, generate temporary handling suggestions for the vehicle desensitized label and feed them back to the corresponding vehicle edge node.

8. The new energy vehicle fault diagnosis method based on cloud-edge collaboration according to claim 7, characterized in that, The process of performing batch defect clustering identification on the fault prediction results uploaded by regional edge nodes and generating batch quality levels corresponding to different batch IDs includes: Based on the fault prediction results corresponding to the vehicle desensitization labels in a certain group, the specific components and fault modes are determined. The batch IDs corresponding to the specific components are extracted from the vehicle desensitization labels. The specific components, the batch IDs corresponding to the specific components, and the fault modes are uploaded to the cloud diagnostic center. The cloud-based diagnostic center obtains the global baseline failure rate corresponding to specific components and failure modes based on all online vehicle data, and obtains the actual failure rate of the batch ID based on the batch ID and failure mode corresponding to the specific component. Based on the actual failure rate of the batch ID and the global baseline failure rate corresponding to the specific component and failure mode, the batch quality level of the batch ID corresponding to the specific component is obtained.

9. The new energy vehicle fault diagnosis method based on cloud-edge collaboration according to claim 8, characterized in that, The process of performing a full-dimensional analysis of health-level fault events uploaded to the cloud diagnostic center and generating health-level fault diagnosis results includes: The system pre-determines battery capacity prediction model and motor life prediction model. It inputs the main abnormal events and the five types of related minor abnormalities and operating conditions in the health-level fault events into the battery capacity prediction model and motor life prediction model, respectively. Based on the battery capacity prediction model and motor life prediction model, it outputs the battery capacity decay trend and the remaining life of the motor, respectively, and feeds them back to the vehicle edge node.

10. A cloud-edge collaborative new energy vehicle fault diagnosis system, used to implement the cloud-edge collaborative new energy vehicle fault diagnosis method according to any one of claims 1 to 9, characterized in that, This includes a cloud-based diagnostic center, which communicates with vehicle-mounted edge nodes and regional edge nodes. The vehicle-mounted edge node is used to collect operating parameters and environmental data of new energy vehicles, and to perform data cleaning, feature extraction and initial fault judgment on the operating parameters. When candidate faults are generated during the initial fault judgment process, all candidate faults are associated and merged to generate one or more fault events. According to the preset static risk value scoring model, the static risk value of each fault event is output, and the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient are obtained. The static risk value is then corrected a second time based on the coupling correction coefficient, degradation trend correction coefficient, operating condition correction coefficient and batch quality correction coefficient to obtain the dynamic risk value. The fault level of the fault event is determined based on the dynamic risk value. Based on the preset regional scheduling strategy, the fatal fault events are assigned to the vehicle-mounted edge node for local execution, the warning fault events are assigned to the regional edge node for execution, and the health-level fault events are uploaded to the cloud diagnostic center for execution. The regional edge node is used to perform single-vehicle early warning level fault diagnosis on the received complete data packet, generate fault prediction results, and perform environmental correlation judgment and false alarm suppression based on the fault prediction results; The cloud-based diagnostic center is used to perform batch defect clustering identification on the fault prediction results uploaded by regional edge nodes, generate batch quality levels corresponding to different batch IDs, and perform full-dimensional analysis on the uploaded health-level fault events to generate health-level fault diagnosis results.

Images